TwoSetAI: Decoding AI Builders

Birju Kadakia grew up in the Chicago suburbs using the local tennis courts and the rec center without knowing he was a community recreation user. He studied engineering at Illinois, was a really early employee at Uber, where he helped launch product teams including Uber Eats, and then led product at The Athletic. He calls himself a product person first. In 2022 he started Rec Technologies to work on what he calls the activity economy: the summer camps, swim lessons, courts, gyms and leagues that, by his account, add up to $400 billion of local spending in the US, with the work in between done on pen and paper.

In this episode, Birju joins TwoSetAI to explain how Rec builds AI agents for parks and recreation. Rec sells to the people who put recreation on, from a city department to the local YMCA to a small pickleball and tennis club, because he says they are the ones bogged down by the work. Angelina brings her own cases: the paper camp catalog that arrives in the mail every year, and the hour she spent at number 77K in an online line for museum tickets that then sold out. He says Rec thinks about that second problem every day, and asks why your agent cannot wait in line for you.

You can learn more about Rec here.

We go deep on how the Refund Agent works, which he demos live on a test city called Pawnee, Indiana. Every class has its own refund policy, which is why the job was hard to automate before. Rec has AI read each policy once and write deterministic code for it, shows the policy back to operators in plain language, and keeps an audit log of the agent. Nothing is inferred at refund time, which he says avoids hallucination altogether. Before a city turns on automated refunds it can turn on refund suggestions, a step he models on the way you would onboard an employee. He then walks through skills and workflows, why a city would use Rec's agents and not hand its policies to Claude, and how the team checks its conversational agents: Langfuse, plus people who take 50 conversations and score them. He also shows the Rec ID, which he calls a recreation passport, and says Rec has over a hundred cities live.

We also get into the harder parts. Asked whether vertical SaaS is dead, Birju says some software is, and names the three things he thinks a platform has to own to be irreplaceable. He gives the accuracy his agents reach task by task and the bar he holds them to, which is to be meaningfully more useful than what you would expect in a world without AI. He explains how Rec got its first data set and its first 10 customers (cold call a hundred, talk to five), how he takes rejection, why every role on his team of about 35 is converging, and why he tells engineers to think about the engineering manager they looked up to. He also names the engineering metric he thinks we will laugh at one day, what he watches instead, and the single word he would teach any founder.

We discuss:

* Growing up on public tennis courts, and why he cannot get excited about a problem unless he can live it

* The activity economy: camps, swim lessons, sports, gyms, and $400 billion of local spending by his count

* A three-sided economy of space owners, operators and all of us, run on pen and paper

* The paper camp catalog, the 8 a.m. sign-up, and why Rec works with the operators first

* An agent that proposes a summer itinerary and tries to book it on booking day ("it's in the works")

* Step one of building Rec: putting schedule data somewhere AI can access

* Number 77K in line: reliability, demand spikes, and the supply side of sold-out tickets

* Why can't you set up your agent to wait in line for you?

* Reservations, memberships, lessons: the software that runs a recreation operator

* Going live: San Francisco's tennis and pickleball courts, and Indianapolis's 800,000 residents

* The Rec ID as a recreation passport

* Live demo of the Refund Agent on Pawnee, Indiana

* Reading each policy once and writing deterministic code, with an audit log

* Skills, triggers and the end-to-end workflow builder Rec is building toward

* Why not pass the policies to Claude? Data, payments and "recreation taste"

* "Some software is dead": the three things he says a platform has to own

* The first data set: one micro problem, pickleball courts in 2022, and 10 cities

* Hallucination as a layered problem: suggestions first, no inference where it matters most

* Accuracy by task, and "meaningfully more useful than what you would expect in a world without AI"

* The first 10 customers in maybe three months, and handling rejection

* Hiring at about 35 people: PMs who ship features, engineers managing many agents, designers who think like PMs

* Why the engineering manager you looked up to is the new baseline for engineers

* The metric he calls a false positive, and what he measures instead

* Speed, Amp It Up by Frank Slootman, and "get off your screen"

Birju Kadakia

* LinkedIn: https://www.linkedin.com/in/birjukadakia/

* Rec: https://www.rec.us/

* Mentioned in the episode: Amp It Up by Frank Slootman, the book he had the whole company read

Timestamps

00:00:00 Why most AI founders are looking in the wrong place

00:01:24 Who is Birju Kadakia and what is Rec Technologies?

00:03:47 What is the activity economy?

00:07:17 Why is a $400B industry still running on pen and paper?

00:10:25 What made you decide to solve this?

00:15:37 How does AI fix the Taylor-Swift-style ticket rush?

00:17:41 Why can't an AI agent wait in line for you?

00:22:28 Why is Indianapolis (800K residents) going live on Rec?

00:24:37 Inside the Rec recreation passport

00:25:58 Live demo: the Refund Agent

00:30:39 How workflows and skills stitch into end-to-end agents

00:33:18 What is "recreation taste" and why does vertical AI need it?

00:36:34 Is vertical SaaS dead? The three pillars that survive AI

00:38:49 How Rec got its first customers

00:39:25 How does Rec avoid AI hallucination on refunds?

00:41:36 What accuracy level do Rec's AI agents actually hit?

00:44:24 How do you handle rejection as a founder?

00:45:42 What roles is Rec hiring for in the AI era?

00:50:38 Why token usage is a false positive for engineer metrics

00:53:01 The one framework Birju would teach every founder

00:56:18 Outro

From leading sales and pricing at SaaS companies in health tech, architecture and agriculture to running her own consulting practice: Jenée Hall is a former chief revenue officer, and founders keep bringing her the same worry, Angelina included. They think they are leaving money on the table. Angelina introduces her as the person who can explain why almost every technical founder who suspects that is probably right, and what to do about it.

In this episode, Jenée joins TwoSetAI for a live pricing workshop on a real early-stage AI agent company: AgensFlow, the startup of Angelina's friend Nicole Königstein, which has no numbers, no packaging and no pricing page yet. Her starting point is that with an AI agent the customer is not buying software, they are hiring a digital employee, so the first question is what job the customer is hiring that agent to do. And pricing, she says, is never a set and forget. It is meant to evolve.

You can find Jenée on LinkedIn here.

We go deep on the layers she uses to map AI agent pricing, what each one charges for, and which she calls the hardest but the highest value. She explains why usage metrics are great internally but not a great value metric (nobody buys electricity because they love kilowatt hours, they buy air conditioning), compares MindBody's monthly fee with a personal trainer software that charges based on the revenue its customers bring in, and lays out a path for a company at this stage: build your costs into a platform fee, get 10 customers paying it, then use what they do with the product to shape tiers and bundles. She also covers how much competitor prices matter, anchor pricing as Steve Jobs used it when he unveiled the iPad, willingness to pay and the Van Westendorp model, and how long a pricing exercise should take, which for a small company she puts at maybe a half day.

We also get into the harder parts: sales teams that discount a price they do not believe in, the founder who was sure of his idea until she sent him to five people in his industry, and a company in Spain whose customers were perfectly happy with the product while it was free. In a rapid-fire round she takes on free tiers, whether to start low or high, and charging what the market will bear, and tells the Monograph story, where by her account testing showed the price could triple. Before the end, Angelina makes a confession about how she set her own prices, and Jenée leaves a reminder that buying anything is an emotional experience, so you still have to appeal to the humans.

We discuss:

* Refereeing youth soccer, and being one of the few women in the room as a sales leader

* Why technical founders treat pricing like a spreadsheet problem

* The costs to count before you set a price, beyond tokens and API calls

* The case study: AgensFlow, an early-stage AI agent company with no pricing page yet

* Why buying an AI agent is closer to hiring a digital employee than buying software

* The layers of AI agent pricing, one at a time

* What a platform fee covers, and why she says the value is somewhere else

* "Nobody buys electricity because they love kilowatt hours": the trouble with pricing by tokens

* Charging for workers in place of seats, and how that mirrors cloud pricing

* Work completed and outcomes: easier to justify, messier to attribute

* MindBody's monthly fee vs. the personal trainer software that charges on revenue

* How much competitor prices matter, from price skimming to penetration pricing

* Steve Jobs, the iPad and anchor pricing

* Why a sales team quietly discounts a price it does not believe in

* Asking customers what they would pay, and the Van Westendorp model

* The five-person pressure test for a founder's idea

* Price before product, or product before price? "It's both"

* The platform fee playbook: 10 paying customers, predictable revenue, then usage data

* How long a pricing exercise takes, from a half day to a three-day workshop

* Rapid fire: free tiers, "tire kickers," and why she would rather have a very low tier of $5 or $8

* Rapid fire: start low and raise, and what you owe legacy customers if you lower

* Rapid fire: charging what the market will bear

* The Monograph story, and why a 50% increase in one year does not sit well

* The psychology of 14.99, and why usage pricing is not the answer just because OpenAI charges by tokens

* Pricing that shapes customer behavior, Angelina's ChatGPT confession, and the book Jenée recommends

Jenée Hall

* LinkedIn: https://www.linkedin.com/in/jenee-arends-hall/

* Mentioned in the episode: Monetizing Innovation by Ramanujam and Tacke, the book she recommends; the Van Westendorp model for willingness to pay; and AgensFlow, the company used as the case study

Timestamps

00:00:00 The pricing lessons every founder needs to hear

00:01:26 Meet Jenée Hall

00:02:03 What Jenée does when she isn't fixing pricing

00:05:10 Why founders treat pricing like a spreadsheet problem

00:08:15 Live workshop: pricing a real early-stage AI agent

00:09:38 The digital employee reframe for AI pricing

00:12:07 The 5-layer AI agent pricing map

00:13:54 Why token-based pricing confuses customers

00:18:01 MindBody vs the personal trainer software (real B2B example)

00:21:11 Do competitor prices actually matter?

00:24:48 Steve Jobs, the iPad, and anchor pricing

00:26:02 Why your sales team silently discounts your price

00:28:07 How to actually talk to customers about pricing

00:30:04 The 5-person pressure test for founders

00:33:00 Should you price your product BEFORE you build it?

00:37:26 The AgensFlow playbook: platform fee, 10 customers, then usage

00:43:43 How long does a real pricing exercise take?

00:46:45 Rapid fire: do I need a free tier?

00:48:17 Rapid fire: price right on day 1 or raise later?

00:53:20 Rapid fire: can I charge what the market will bear?

00:54:54 The Monograph story: from $29 to triple

00:59:02 Why pricing is strategy, not accounting

01:02:06 The pricing insight that changes user behavior

01:04:31 I asked ChatGPT how much to charge: a confession

01:05:39 Book recommendation: Monetizing Innovation

01:06:31 Outro

From naive Bayes on CPUs to GPT-1 and GPT-2 around 2019 to a book on production RAG: Ofer Mendelevitch says he has seen a lot of cycles. He is a co-author of O'Reilly's Hands-On RAG for Production and an independent AI advisor. He started the book a little more than a year and a half ago, drawing on what he learned helping customers implement RAG at Vectara and watching open source projects such as LangChain and LlamaIndex, and he wrote it for the point where a prototype has to become a production system.

In this episode, Ofer joins TwoSetAI to take on the claim that RAG is dead. By his count it has come around every two months for the last two years, mostly from people trying to market something new, and his answer is that the R stands for retrieval, which is still extremely important even with agents. To show what that means, he takes a real use case of Angelina's, checking whether the references and numbers in AI-generated text hold up against their sources, and designs the system with her step by step.

Ofer's book, Hands-On RAG for Production, is published by O'Reilly. You can find more of his writing here.

We go deep on the distance between the demo and production. Ofer explains why semantic search can miss the difference between 85 and 95, where hybrid search with BM25 and TF-IDF earns its place, and why a one-shot query gives way to an agent that calls retrieval as a tool, 10 times if it has to. He covers tables, images, audio and video, including what he calls the red button problem and his read on where video embedding models stand today. He describes an agent harness as the same thing as an agent with a better implementation, and says a knowledge graph is worth building only after you think through the ROI: not, in his example, if it solves 0.1% of your queries. On cost he points to the first ingest, where he says real documents can run to 15,000 pages, and on evaluation to UMBRELA, a method from Jimmy Lin's lab at the University of Waterloo in which an LLM scores each retrieved chunk from zero to three, with no golden answer required.

We also get into the harder parts: when to build, when to buy RAG as a service, and when a vertical tool falls short because the feature you need today is 3 years out; developer relations as a growth strategy for startups, how he would budget and measure it, and why, as a big user of AI himself, he tells teams to keep their own voice in what they publish; and the layoffs at large companies, which he calls a bit of an overshoot while recognizing the pain for real people with real families. Then the conversation turns to careers. Ofer says where he thinks engineers and data scientists end up once coding agents write the code, names the risk that concerns him most, and recounts what he told a high school computer science club, all of them just accepted to computer science programs, when they asked whether they had made a terrible mistake.

We discuss:

* Naive Bayes, a world with no GPUs, and using GPT-1 and GPT-2 around 2019

* Why he wrote Hands-On RAG for Production, and what the "I can talk to my PDF" demo leaves out

* "RAG is dead" claims every two months, and why the R still matters with agents

* Angelina's use case: checking AI-generated references and numbers against the source material

* Why semantic search can miss the difference between 85 and 95

* Hybrid search: BM25 and TF-IDF for product SKUs, codes and exact words

* One-shot RAG vs. an agent that calls retrieval as a tool

* Tables as first-class citizens, and why chunking loses the column names

* Images, audio and video: the red button problem, and where video embedding models stand

* Harness engineering: "a better agent"

* Knowledge graphs, multihop questions and thinking through the ROI

* Keeping an index fresh: document hashes, triggers, push and pull

* Where the cost sits: the first ingest, 15,000-page PDFs, API calls vs. self-hosted open models

* Quality as a cost: 50 engineers, 6 months, and users who say "I don't care"

* Evaluation without a golden answer: UMBRELA and the work from Jimmy Lin's lab at the University of Waterloo

* Build it yourself, buy RAG as a service, or use a vertical tool

* What DevRel is for: teaching the product and bringing feedback back to the company

* Product-led growth vs. sales-led growth, and what hackathons, events and blog posts are for

* AI slop, and keeping your own voice

* Budgeting and measuring DevRel without over-engineering it

* Layoffs as an overshoot, and where engineers and data scientists end up once coding agents write the code

* What he tells his own kids, and what concerns him as a society

* The high school seniors who asked whether choosing computer science was a mistake

* Why "solve Fibonacci in five lines of Python" stops working as an interview test

Ofer Mendelevitch

* LinkedIn: https://www.linkedin.com/in/ofermend/

* Medium: https://ofermend.medium.com

* Book: Hands-On RAG for Production (O'Reilly)

* Vectara: https://www.vectara.com

* Mentioned in the episode: UMBRELA and the reference-free evaluation work from Jimmy Lin's lab at the University of Waterloo; LangChain and LlamaIndex

Timestamps

00:00:00 Is RAG actually dead?

00:04:08 Why write a RAG book when everyone says RAG is dead?

00:06:24 Can you build a hallucination checker with RAG?

00:10:26 Why does AI read 85% and 95% as the same number?

00:14:13 Why do agents beat one-shot RAG on hard questions?

00:20:06 How do you search tables, images, and video?

00:28:05 When is a knowledge graph actually worth it?

00:33:09 Which part of a RAG system costs the most?

00:39:21 How do you grade AI when there's no answer key?

00:45:59 Should you build your own RAG system or buy one?

00:49:41 Is developer relations worth it for a startup?

01:01:17 Will AI replace engineers and data scientists?

01:06:22 What should you study if AI can already code?

01:09:04 Outro

From a head of engineering to a company of one: Ali Parandeh was on TwoSetAI before, talking about vibe coding and building production AI systems. Since then he has left his job and started his own company, Build Your AI. He is the author of the O'Reilly book Building Generative AI Services with FastAPI. At the time of this conversation he is a solo founder who teaches engineers AI and runs his own work on a handful of AI tools.

In this episode, the first half of a longer conversation, Ali joins TwoSetAI to talk about OpenClaw, the open-source agent framework with 250k GitHub stars that people are leaving to run overnight. He calls these extremely exciting times: a friend of his talks to his own agent on WhatsApp and uses it as a fitness trainer. But Ali is not running it himself. He wants privacy and safety around his data, and he points to someone who racked up $800 or $300 of costs because their agent ran into a loop and couldn't decide to stop. Angelina then asks the question in the title: when will anyone be able to hire a reliable AI employee? Ali gives his read on where the technology stands.

You can learn more about Build Your AI here.

We go deep on the automation stack of a solo founder. Ali's rule is that you don't need hundreds of tools, you need a few cornerstone tools: one general assistant such as Claude or ChatGPT, a way to talk to all your notes and documents, a transcriber, and accounting software. His own setup runs on Notion, his transcriber and knowledge management system, with custom Copilots inside Microsoft 365 for sensitive information and Figma Make for prototypes. He walks through how he rebuilt his own website as a pipeline: two or three days on the copy with Notion AI, around 100 hero section designs from Relume, and JetBrains AI for the animations, with Ali quality controlling the outputs between each stage. He also tells the story of a teacher who, by his account, built a whole working app in Lovable and then had no idea how to deploy it into Google Cloud or Microsoft Azure, which the school wanted.

We also get into the harder parts: what heavy use of AI does to the person using it. Ali says calculators have made all of us really bad at arithmetic in our heads, and that ChatGPT is like that for language. He describes an MIT study in which, as he tells it, three groups wrote an essay with ChatGPT, with web search, or with pen and paper, and most of the brain of the ChatGPT group was kind of deactivated. His own rule is to own any output that AI generates: AI will not do 100% of the job, he tells the engineers he teaches, it will do 80%. Overuse, in his words, is like eating junk food. He is not arguing against AI, which he teaches and uses for a lot of his work; what he wants is balance and moderation. He also explains why his worry is a random business using AI in safety-critical applications and not the singularity, why you do not want ChatGPT as a therapist, and how he tested job candidates he assumed would use AI. The episode closes with Angelina laying out her own three layers of automation and Ali's advice on designing a workflow before building it.

We discuss:

* OpenClaw at 250k GitHub stars, and people letting AI run overnight

* The friend who uses his own agent on WhatsApp as a fitness trainer

* Why Ali is not running OpenClaw himself: guardrails, privacy and API costs

* The $800 or $300 bill from an agent that ran into a loop

* When we will be able to hire a reliable AI employee, and marketplaces for AI employees

* Custom agents in Notion, and where he is more confident putting sensitive data

* "You don't need hundreds of tools": the few cornerstone tools a solo founder needs

* Notion as transcriber, knowledge management system and CRM, and what he pays for Notion AI

* Custom Copilots in Microsoft 365 for sensitive information

* Switching between Gemini, Claude and ChatGPT for different jobs

* Figma, FigJam and Figma Make for slides, workshops and prototypes

* The teacher who built a full app in Lovable and could not deploy it for a school

* How he rebuilt his website with Notion AI, Relume and JetBrains AI

* Calculators, arithmetic in our heads, and ChatGPT

* Learning video editing by hand in CapCut: 3 hours for a 10-minute video

* The quiz he built with Claude on a Daniel Priestley framework

* The MIT study he describes on brain activity when writing with ChatGPT

* "Own any output that AI generates", and the kitchen analogy

* Why overuse is like eating junk food

* The risk he worries about: AI in safety-critical applications

* ChatGPT as a therapist, homework, and how he tested job candidates who use AI

* Angelina's three layers of automation, and his advice on triggers, n8n and Mermaid diagrams

Ali Parandeh

* Build Your AI: https://buildyourai.com

* His book: Building Generative AI Services with FastAPI (O'Reilly)

* Mentioned in the episode: Daniel Priestley, Scorecard and the quiz framework Ali learned from him; the MIT study on brain activity and AI use, as Ali describes it; OpenClaw, Notion, Figma Make, Relume, JetBrains AI, CapCut, Lovable, n8n and make.com

Timestamps

00:00:00 People are letting AI run overnight. Should you?

00:00:42 OpenClaw hit 250K stars. Why won't he use it?

00:08:26 When will we actually hire reliable AI employees?

00:11:15 What does a solo founder's real automation stack look like?

00:21:59 She built a full app with AI, then couldn't sell it to schools

00:24:02 The 4-tool pipeline behind his entire website

00:26:27 "Calculators made us all bad at math"

00:31:38 The MIT brain study every AI user should know

00:37:18 Is AI dependency the new junk food?

00:42:29 His girlfriend is a therapist, and she hates ChatGPT therapy

00:44:49 How should Angelina build her automation stack?

00:50:52 Outro

From two startups that kept building the same integrations to the company that sells them: Gil Feig is the co-founder and CTO of Merge. He and his co-founder Shensi met the problem from different sides, one in cybersecurity connecting to ticketing systems, the other in recruiting tech connecting to applicant tracking systems, and the company started in 2020. It began as integrations for SaaS products. Then AI agents arrived, and by his account Merge is now the context provider for dozens of Fortune 500s.

In this episode, Gil joins TwoSetAI to argue that the bottleneck for AI agents is the data underneath them, and that a smarter model would not fix it. Put the smartest person on the planet in front of the question "Which of my customers were upset last year?" and it does not matter how smart they are, he says: they still need the data to go through. An agent has the same problem, he adds, because the API itself does not make it easy to get all of that data out in real time.

You can learn more about Merge here.

We go deep on how agents should reach company data. Gil walks through Merge's two products, Merge Unified for synced data and Agent Handler for live, MCP-style connections, and when each one fits: a live lookup for ticket number 347, a synced copy for questions that span the whole data set. He explains why handing an agent API keys and stuffing every result into the context window breaks down, why he says the vast majority of public MCP servers are unusable and Merge rebuilds all of them, and why RAG is far from dead, only no longer the one option. He also covers evals, where an agent judges whether an answer means the right thing whatever its wording, what a context graph is and how close it sits to a knowledge graph, and why Merge syncs permissions twice, in a second engine built by someone who did not see the original code base, so that a new grad cannot see the board deck.

We also get into the harder parts: why he told his whole company not to bother with OpenClaw unless they are sure of what they are doing, why an agent holding an admin API key can send it anywhere, and the logo ladder an infrastructure provider climbs before it reaches the enterprise. Then the conversation turns to people. Gil says AI is already a better coder than most engineers he knows, compares enterprises to the ship that everyone could see was going to hit the bridge in Baltimore, and still says Merge wants to hire more right now. He closes with the book he recommends, The Singularity Is Near, his read on the year it names for AI reaching human intelligence, and what the advice "make all your money now" says about the fear of a post-AGI world.

We discuss:

* What Merge does, explained for an elementary schooler, a developer and a C-suite

* How two startups with the same integration problem led to Merge in 2020

* From SaaS integrations to "the context layer"

* Merge Unified and Agent Handler: synced data and live connections

* "Which of my customers were upset last year?" and why MCP cannot answer it in real time

* The naive agent architecture: API keys, too many tools, everything in the context window

* Why the smartest person on the planet could not answer without the data

* MCP in plain terms: scoped tools in place of an API key

* OpenClaw: why he told his whole company not to bother with it right now, and where to run it if you do

* Mac Minis selling out, and the old company devices his employees now buy

* Why he says most public MCP servers are not enterprise grade, and the vast majority are unusable

* "RAG is not dead. It's far from dead": when a synced copy beats a live lookup

* Evals: testing answers that are worded differently every time

* Hallucination, LangChain and CrewAI, and what he says changed since four months ago

* What a context graph is, and how it differs from a knowledge graph

* A walkthrough of the Agent Handler dashboard: connectors, tool packs and the playground

* The hardest part: the sheer volume and breadth of data, and integrations that constantly break

* Why an agent with an API key could post it to a hacker's URL

* Syncing permissions twice so a new grad cannot see the board deck

* Climbing the logo ladder, and hiring learners and growers

* "You won't need software engineers," the bridge in Baltimore, and how much time is left

* Why Merge wants to hire more anyway

* Where a non-technical person should start, and why a Claude skill is a text file

* What Gil automates himself, how Merge differs from Glean, and the internal agent Netflix built on Merge

* The Singularity Is Near, the year it names, and "make all your money now"

Gil Feig

* LinkedIn: https://www.linkedin.com/in/gilfeig/

* Merge: https://merge.dev

* Mentioned in the episode: The Singularity Is Near, the book he recommends

Timestamps

00:00:00 A CTO's prediction every software engineer should hear

00:01:42 What does Merge actually do?

00:04:34 Why two friends quit during COVID to build this

00:08:29 The naive AI agent architecture that always breaks

00:11:43 Would the smartest person alive solve this?

00:14:25 Why he told his whole company not to use OpenClaw

00:16:45 Why most MCP servers are unusable

00:18:36 Is RAG actually dead?

00:21:23 How do you know your AI integrations actually work?

00:24:30 What is a context graph?

00:29:52 The hardest part of connecting everything

00:33:18 Why Merge rebuilds every integration twice

00:35:07 How to climb the enterprise logo ladder

00:36:13 AI is already a better coder than most engineers

00:41:42 What AI automations does a CTO actually use?

00:43:20 How is Merge different from Glean?

00:45:24 The Singularity Is Near, and the predictions are landing

00:47:05 "Make all your money now"

00:49:00 Outro

From coding algorithms on paper in her computer science exams in Switzerland to writing the O'Reilly books on transformers and AI agents: Nicole Königstein is an author, researcher and consultant whose PhD research is on self-improving AI agents. She calls herself a systems thinker, and says what she cares about is helping people understand the domain in a deeper sense.

In this episode, Nicole joins TwoSetAI to take the hype out of AI agents. Asked whether an agentic system is a bunch of if-then rules with prompts, she answers with the picture she uses at the beginning of her book, and she is plain that today's agents are not fully autonomous: people still give them the vocabulary, the tools and the rules. A system that can alter its own code and architecture should be technically feasible at some point, she says. Her question is whether we really want it.

You can find Nicole's O'Reilly books, including AI Agents: The Definitive Guide, here.

We go deep on how to design a multi-agent system, with a patent filing system as the worked example: start from the roles and tasks involved, decide what has to be verified, and choose a topology, from a tech lead with sub-agents to independent minds around a shared blackboard. Nicole explains why lining up more agents will not fix a system that is not designed properly, pointing to a DeepMind paper and her own work, and why every additional layer, verifier and call consumes tokens. She covers evaluating as early as possible with subject matter experts, the open-source RULER framework she uses as a relative judge, and reinforcement learning that updates a model's weights so it behaves better in open-ended tasks with no single correct answer. From her own comparisons she reports that the more expensive models overthought their tasks, in some cases spending four times as many tokens as a smaller open-source model with results that were not as good. She also names AgensFlow, the tool she was developing to choose models for a task, which she had not yet released at the time of recording.

We also get into the harder parts: thought experiments about an AI given the wrong incentive, why she believes hallucinations cannot be fully eliminated, and what happens when someone vibe codes a multi-agent system without knowing how to design one. It will pass the tests, she says, and after the first months in production the CTO will be stomping into your office because you burnt lots of tokens. Her line on coding agents is that you cannot fully outsource the thinking. She closes on who AI agents will replace, how her university had students code on paper and work with real clients, and why a junior entering the job market should enjoy the process of learning.

We discuss:

* The version of Nicole beyond the book: a systems thinker

* How she would explain an AI agent to a non-technical person, starting with transformers and tokens

* Why AI feels so human: attention, pattern recognition and memory

* The stateless agent that could not remember her name

* Are agents a bunch of if-then rules with prompts? Her answer

* Topologies as team structures: a tech lead with sub-agents, or a room with a whiteboard

* Why she says today's agents are not fully autonomous

* Her PhD research on self-improving agents, and what reinforcement looks like in people

* "Do we really want this?": wrong incentives, thought experiments and guardrails

* The MIT experiment on autonomous driving, and why human goals are hard to write into a system

* Orchestrated autonomy, and the OpenClaw post about a canceled workshop

* Designing a patent filing system: roles, tasks, verification and a swarm of researchers

* Why more agents will not fix it, and when one good model with specific guidance is enough

* Evaluating early, bringing in subject matter experts, and stress testing with angry users

* Every additional layer consumes tokens, and RULER as a relative judge

* Which agent needs to know what: memory and context across a system

* Training with reinforcement learning: updating weights for behavior, not knowledge

* Why she believes hallucinations cannot be fully eliminated, and checking sources across model families

* Long context windows and the models that got confused

* Comparing models as judges, and the expensive ones that overthink

* AgensFlow, the tool she is building to choose models, and where the name comes from

* Skill files as a contract between employer and employee: treat the agent like an intern

* Vibe coding a multi-agent system: it passes the tests, then the tokens burn

* Who agents will replace, and how juniors build judgment: coding on paper and real client projects

* Ada Lovelace, Feynman, enjoying the process, and a paragraph from chapter six of her book

Nicole Königstein

* LinkedIn: https://www.linkedin.com/in/nicole-koenigstein/

* O'Reilly books: https://www.oreilly.com/search/?q=nicole+koenigstein

* AgensFlow (open source): https://github.com/Nicolepcx/AgensFlow

* AgensFlow paper: https://arxiv.org/abs/2605.27466

* Mentioned in the episode: Surely You're Joking, Mr. Feynman; Ada Lovelace; RULER, the open-source relative judge framework; and a DeepMind paper on the science of agent systems

Timestamps

00:00:00 The woman writing the AI agents textbook: what's she like off stage?

00:02:37 How would you explain AI agents to a non-technical person?

00:09:51 Are AI agents just if-then rules with better prompts?

00:16:26 Her PhD is secretly about AI that improves itself

00:19:43 The thought experiment that kills "just automate everything"

00:27:17 Building a patent filing system from scratch

00:33:23 Why adding more agents won't fix your system

00:38:47 Every additional layer burns tokens

00:47:27 Can hallucination ever be fully solved?

00:53:44 Why expensive AI models overthink and produce worse results

00:55:37 She built a tool that picks the right model

00:57:45 Skill files as contracts: governance without jargon

01:01:09 What happens when you vibe code production?

01:05:18 Will AI agents replace software engineers?

01:13:27 Ungoverned execution isn't recoverable

01:14:24 Outro

From starting college in computer science and working in VR a decade ago to six years of investing at Lightspeed and now a venture firm of her own: Mercedes Bent is a partner at Premise Ventures, the firm she started with a co-founder. She describes it as generalist, investing in everything from research models to B2B dev tools to the app layer and fintech, and she says she loves investing in consumer because it is so hard.

In this episode, Mercedes joins TwoSetAI to answer a builder who watched every YC video and took away one rule: do B2B, do the boring business, do not try consumer. She does not think that advice is entirely wrong. It is easier, she says, to get a B2B company from seed to Series A. But she cites a report that only 6% of VC funding went to consumer startups in 2024, which in her reading means far more venture-backed competition on the B2B side, and she explains why she does not expect OpenAI to eat everything in consumer.

You can learn more about Premise Ventures here.

We go deep on where startup ideas come from. Borrowing another founder's analogy, Mercedes says a new digital product has to be built on new land, and that new land appears when a recent tailwind moves through a whole population or economy. She walks through the sources of tailwinds she looks at and the two consumer behavior changes she sees under many digital products today: voice, and a younger generation's approach to money that she traces back to the affordability of home ownership. She also lays out how she evaluates a company, founder first, with a seven-point framework for founders, then product, market and the deal, and says the skill of an investor is telling the top 0.01% from the top 10%.

We also get into the harder parts: founders who get more enamored with the solution than the problem, which she saw when she worked in VR, and the advice she repeats to technical founders: just go talk to the customers. In a rapid-fire round on what founders believe about VCs she covers decks, warm intros, Ivy League pedigree and why going viral on GitHub counts for more than going viral on LinkedIn or Twitter. She is direct about the numbers for women: by her account 2% of the CEOs of VC-backed companies are women, and her advice is to build your brand and show up twice as good all the time. She closes on why every AI product looks like a chatbot, the user dials and voice buttons she wants to see, and the seven ways she thinks the AI bubble could get popped, of which she has time to name only a few.

We discuss:

* The YC advice Angelina took to heart: do B2B, do the boring business, do not try consumer

* Why B2B is easier from seed to Series A, and why consumer gives you minutes, not a year, to show value

* Why she does not invest in consumer products before launch

* The report that only 6% of VC funding went to consumer startups in 2024

* Will ChatGPT kill every consumer startup? What Google's history suggests

* The business models she says always work: networks, marketplaces, platforms and systems of record

* Voice as one of the biggest changes in how consumers use digital platforms

* Building on "new land"

* Home ownership, riskier assets and a generation's changed approach to money

* What she sees in spending, saving and debt

* The sources of tailwinds, and why it has to be something that could not have been done five years ago

* What a VC looks at beyond pedigree, demo and market size: founder, product, market and the deal

* Her seven-point framework for founders, and the top 0.01% against the top 10%

* Why the market for black cars looked tiny when Uber started

* The founder as an influential communicator of a vision 10 or 15 years out

* The sci-fi tech club she runs so that part of her brain lives in the future

* Getting more enamored with the solution than the problem: what VR taught her

* "Just go talk to the customers"

* Running her communities with agents, and why she says investing is inherently a people-based role

* Why relationships win when the best companies have too much capital coming their way

* Rapid fire: decks, team over idea, co-founders, Ivy League pedigree, age and warm intros

* Going viral on GitHub against going viral on LinkedIn or Twitter

* Women founders: the 2% and 20% numbers, building a brand, and showing up twice as good

* Why every AI product looks like a chatbot, and what early web pages have to do with it

* User dials, voice buttons, agentic websites, OpenClaw, and the seven ways the AI bubble could get popped

Mercedes Bent

* LinkedIn: https://www.linkedin.com/in/mercedesbent/

* Website: https://mercedesbent.co

* Premise Ventures: https://premise.vc

* Mentioned in the episode: her article on voice buttons; and the Sci-Fi Tech Club and Surreal, two communities she runs

Timestamps

00:00:00 "We're so dumb for starting a VC firm as two women"

00:00:51 Why does YC tell founders to avoid consumer?

00:01:39 Is consumer AI actually harder than B2B?

00:04:36 Is consumer less competitive than enterprise?

00:06:28 Will ChatGPT just kill every consumer startup?

00:08:27 Why is voice the biggest behavior shift in 2026?

00:09:31 How do builders find "new land" to build on?

00:10:24 How did homeownership reshape a generation's risk appetite?

00:13:06 Where are spending and savings trending now?

00:13:45 What are the 5 sources of startup tailwinds?

00:17:20 What does a VC actually look for beyond the pitch deck?

00:18:28 The 7-point founder framework that filters 99%

00:21:47 Why founders need to be Olympians, not just athletes

00:23:17 Inside the Sci-Fi Tech Club: training your brain for the future

00:24:59 When founders fall in love with the solution instead of the problem

00:26:28 Why "just go talk to customers" is still the most ignored advice

00:28:16 Building an agentic VC firm with two people

00:30:21 When VCs compete for founders instead of the other way around

00:33:56 Rapid fire: myths founders believe about VCs

00:37:25 The real data on female founders and funding

00:41:23 Why is every AI product a chatbot?

00:43:46 What new AI interfaces could actually look like

00:47:35 Is the AI bubble about to pop?

00:48:46 Outro

Wiley Jones comes to enterprise software from electrical engineering, mechanical engineering and control systems. He is the CEO of Doss, which had just raised $55 million when this was recorded. Doss builds what it calls ARP for physical operations companies, many of them in the mid-market: businesses with several sales channels, warehouses and product lines that, he says, cannot work their way to the answer one Zapier workflow at a time. He describes ARP as the evolution of what ERP was always meant to be, and says an implementation usually takes about 3 to 4 months.

In this episode, Wiley joins TwoSetAI to make the case for self-evolving software. Software engineers already point these systems at their own codebases, he says, so why does the application itself just sit there? His answer is to put the application and its life cycle in the same runtime, so the system can inspect itself, suggest an improvement, test it on a branch of itself, and flag a human when it touches a hot path. People steer with goals, heuristics and rules, and can enter the loop at any point. He compares the job to gardening.

You can learn more about Doss here.

We go deep on what AI native means to him: the intelligence sits inside the system and has agency, the way someone fluent in Spanish does not pull out a phone and start Google translating. At Doss that rests on an intermediate representation of the software, something between a README and the codebase. It is graph-like, it holds the schema, the business logic and the interfaces (nouns, verbs and adjectives, in his words), it is written in an internal domain specific language called ZSL, and it lives in a standard Git life cycle. He calls Git the most underappreciated creation of the last 50 years. From there he gets to hallucination, which he calls an interesting word, and gives his own account of what it is. He argues that software engineers are re-deriving from first principles something his own field worked out long ago, and explains where an error matters, where it does not, and how a system can be set up to correct itself.

We also get into the harder parts. Wiley calls SaaS a blip, a window when very simple software could extract a lot of value, and says great products built on very thin technology will be wiped out unless they rebuild or become a brand people love. He credits the enterprise builders who came before, naming Workday, PeopleSoft and Salesforce, with composable systems that are really hard to build. The rest of SaaS, he thinks, is not doomed but will be repriced against the cost of intelligence per token. He gives his test for build versus buy (what is the cost of an error, and is it even worth building), explains why agents will not speed up five people getting on the same page, and describes the shared company brain he calls the single most important thing Doss has built for its own productivity. On hiring, he says he does not care whether a candidate knows PyTorch, and he closes on what he thinks is still durable.

We discuss:

* What ARP is, and how it relates to ERP

* Value engineering: a roadmap exercise before and after the contract is signed

* What he thinks "forward deployed engineering" is often a code word for

* Who needs this: physical operations companies with messy, real-world complexity

* Why mid-market companies buy and do not build, and what a long implementation costs a fast-changing business

* A demo in a couple of hours vs. an implementation of 3 to 4 months, and where that time goes

* AI native: the fluent speaker who does not need Google Translate

* Context engineering as one half of it, agency as the other

* The intermediate representation between a README and the codebase

* Schema, business logic and interfaces as nouns, verbs and adjectives

* Why your Notion does not evolve itself, and what would have to change

* Humans as observers who can enter the loop, and the "hot path" rule

* Control theory, and what software engineers are re-deriving from first principles

* Hallucination as a selection error, and bounding the output into discrete outputs

* The sandwich theory of determinism: concrete inputs, concrete outputs, and a middle optimized for speed and cost

* Why agents will not speed up human coordination

* ZSL, the internal DSL, and Git as the most underappreciated creation of the last 50 years

* Why he says pre-SaaS enterprise builders already thought this way

* Great product, thin technology: who gets wiped out, and the brand option

* Build vs. buy: payroll, Ramp, a project tool built in a couple of days, and why not to vibe code a Calendly

* SaaS "repriced" against the cost of intelligence per token

* "More to build now than there ever has been in the history of history"

* Finding customers the Jane Goodall way: a food and beverage conference, not an ERP conference

* His daily AI routine, and the shared company brain with agents as librarians

* Hiring for creativity, agency and judgment, the red wire and green wire problem, and what stays durable

Wiley Jones

* LinkedIn: https://www.linkedin.com/in/wileycwjones

* Doss: https://www.doss.com/

* Mentioned in the episode: Softwar, a book about Oracle in the early 2000s; and Walter Isaacson's book on the history of computing

Timestamps

00:00:00 "Self-evolving software sounds like science fiction"

00:01:17 The hardware engineer rebuilding enterprise software

00:01:32 Why is your supply chain software broken?

00:04:00 What does a value engineer actually do?

00:09:24 Stuck between spreadsheets and SAP?

00:13:00 Why do enterprise systems take 2 years to install?

00:16:05 Why does a demo take 2 hours but implementation takes 3 months?

00:17:55 What does AI-native actually mean?

00:23:20 The language between code and English

00:27:42 Self-evolving software: why Notion doesn't evolve itself

00:32:03 Control theory meets AI agents (200 years of forgotten math)

00:32:59 Is hallucination just a selection error?

00:38:50 Why invent your own programming language?

00:41:06 Why pre-SaaS enterprise builders already thought this way

00:44:18 Should companies rewrite their products for AI?

00:46:42 Build vs buy in 2026: how do you decide?

00:49:28 "More to build now than in the history of history"

00:51:10 B2B customer acquisition the Jane Goodall way

00:53:38 How does a CEO use AI every day?

00:55:40 What if your whole company shared one brain?

00:58:14 Why he doesn't care if you know PyTorch

01:00:56 The bomb defusal thought experiment

01:03:08 The book that changed how he thinks about purpose

01:05:32 What's still durable?

01:07:38 Outro

From a high-frequency trading company to early-stage startups in New York to building AI for finance: Deepak Bapat is the CTO and co-founder of Tabs. As he explains it, Tabs starts by reading a company's contracts and then works on two tracks: making invoicing, billing and payment efficient with as little revenue leakage as possible, and getting the company through its next revenue audit. He says part of the reason the company exists is a shortage of accountants.

In this episode, Deepak joins TwoSetAI to explain what it takes to build AI that a finance team will trust with its revenue. Angelina puts it to the test. She describes how she would build his contract processing feature herself, with a PDF, Claude, a database and a cron job, and he walks through what that version misses, including the line he draws between the work he hands to a language model and the work he will not.

You can learn more about Tabs here.

We go deep on the pipeline behind contract processing: classifying a document with embeddings before a language model reads it, the context objects Tabs generates during implementation, and his numbers on accuracy. By his account generic models without the right context degrade into the mid-80s on complicated contracts, while the right context keeps extraction well into the high 90s. Deepak describes the anomaly detection that flags a contract that fits no known bucket, the human in the loop that Tabs requires when it launches new features, and the commercial graph: patterns across many merchants and their customers, such as what it takes to get an invoice through a large company's AP portal, which he argues a standard horizontal AI product cannot get. He also says some earlier models do a better job on contracts than newer ones given the same context object, and explains why he wants to stay on simpler, cheaper models.

We also get into the harder parts: why he treats finance as a special case for replacing people with agents, since someone still has to sign their name to the numbers, and the CFO design partner who told him the whole reason they had hired him was to have someone to blame when things go wrong; his prediction that a company will grow to 100 million in ARR with one person running finance across a stack of tools; why Tabs made an early bet against a chat interface and does not let users prompt on top of a context object; and hallucinations, which he builds on the assumption will happen. On the engineering side, he admits how quickly his own coding skills deteriorated, explains why he still invests in junior engineers and spends 20 minutes every day with an engineer on the team, and says one of Tabs' coding exercises became a code review exercise.

We discuss:

* Explaining Tabs three ways: to an elementary schooler, a finance manager on spreadsheets and a CFO

* The contract as the start of a financial relationship, and the two tracks: revenue leakage and the audit

* Who uses it, and the accountant shortage behind the company

* The moment a finance team stops being skeptical

* AI agents for reconciliations and journal entries, and what sits behind the month-end close

* Billing as a customer relationship problem

* Build vs. buy, from an engineer who worked on an in-house billing platform 12 years ago

* The commercial graph: what many merchants and customers show that one finance career cannot

* AP portals where it can take years to get an invoice paid

* Anonymizing and aggregating data, and why Tabs does not train its own models

* Why finance is a special case for agents: liability, and signing your name

* "The whole reason they've hired me is to blame someone when things go wrong"

* One person running finance at a company with 100 million in ARR

* Orchestration vs. choreography between agents from different systems

* Angelina's build with Claude and a cron job, and what it misses

* The pipeline: embeddings to classify, LLMs with context objects to extract, deterministic workflows for the invoices

* Mid-80s vs. high 90s: what the right context does to contract accuracy

* Anomaly detection, and context objects that recalibrate from a merchant's changes

* Why some earlier models beat newer ones with the same context object

* Hallucinations, and keeping a human in the loop

* The bet against a chat interface, MCP and background agents

* Why Tabs does not let users prompt on top of a context object

* An engineering team of around 40 to 45, token costs, and doing more with more

* Coding skills that deteriorate, investing in junior engineers, and a code review interview

* The book he calls his favorite of all time

Deepak Bapat

* LinkedIn: https://www.linkedin.com/in/deepakbapat

* Tabs: https://www.tabs.com/

* Mentioned in the episode: The Age of Wonder by Richard Holmes, the book he recommends; and a paper he cites on GPT 5.2 failing to count to five

Timestamps

00:00:00 What if AI could run your entire finance department?

00:01:08 How a high-frequency trader ended up building AI for finance

00:03:25 What does an AI billing platform actually do?

00:09:15 Are CFOs already using AI agents for reconciliations?

00:11:05 Should you build billing in-house or buy a platform?

00:13:36 What is a commercial graph and why can't horizontal AI build one?

00:17:54 Can AI agents actually replace human finance teams?

00:24:21 Will one person run finance for a $100M company?

00:27:07 How will AI agents from different platforms communicate?

00:28:54 Can you build an AI billing product with Claude and a cron job?

00:35:23 Is AI accuracy a model problem or a context problem?

00:39:48 Why did this CTO never build a chat interface?

00:45:10 Are developers losing their coding skills because of AI?

00:46:42 Why invest in junior engineers when AI ships faster?

00:48:36 Should coding interviews be replaced with code review?

00:51:16 The book every interdisciplinary builder should read

00:53:08 Outro

From a sensor he developed in grad school and turned into his first company, to a voice agent startup that failed, to a company he calls a rocket ship: Tom Shapland works on LiveKit Inference at LiveKit, which Angelina introduces as the open-source platform behind OpenAI's voice mode, powering voice agents for everyone from startups to 911 operators. At the time of this conversation, Tom says, LiveKit had just raised its Series C and become a unicorn, with around 100 people. He did not start coding until grad school, and he says his happy place is being absorbed in a problem.

In this episode, Tom joins TwoSetAI to take apart the voice agent most developers would build first. Angelina describes hers, a sales training coach made of Whisper, Claude, ElevenLabs, a WebSocket and a voice activity detector, and Tom calls it a great first pass that will work but come out brittle and slow. From there he walks through the models a voice agent needs, why LiveKit moves audio over WebRTC, and how to shorten the pause before a reply: once it passes about a second and a half, he says, users feel the agent is taking too long.

You can learn more about LiveKit here.

We go deep on the choices behind a production voice agent: why, by his account, most agents that generate revenue run a pipeline of speech-to-text, an LLM and text-to-speech and not a speech-to-speech model; how preemptive generation starts the LLM before the user has finished speaking; and what he calls the biggest problem in voice agents, which Angelina's daughter ran into when she ended up arguing with ChatGPT's voice mode. He also covers LiveKit Inference, which puts the voice models behind one key and one contract; why a coding agent working from stale training data is "the blind leading the blind"; and real-time avatars, where he says people talk a lot more to a face than to a waveform. One user, he says, told him a problem was affecting 1% of their calls: "So that's like 100,000 minutes yesterday."

We also get into the harder parts: hallucinations in regulated industries, and his answer of breaking one giant prompt into many small agents; what the moat is when anyone can fork the code; and whether coding agents change the gap between junior and senior engineers. Tom says the toothpaste is not going back in the tube, yet he still has to babysit his own coding agent. On the founder side he compares his two companies, says the next thing he builds would have to be weird and something he feels needs to exist, and admits he never wants to do B2B sales again. He closes on why he thinks judgment is emotional, not intellectual, and on the book his 20-year-old self would tell him to read again.

We discuss:

* How Tom explains LiveKit to his kids, who are eight and 10

* The two parts of LiveKit: a WebRTC transport layer and an open-source agents framework

* Angelina's first-pass stack for a sales training agent, and why it comes out brittle

* WebRTC vs. WebSockets: dropping packets to keep the conversation in real time

* Preemptive generation, and the second and a half before a reply feels late

* Why most revenue-generating voice agents use a pipeline, not speech-to-speech

* Consumer apps vs. B2B workflows such as lead qualification

* Angelina's daughter arguing with ChatGPT's voice mode

* What Tom calls the biggest problem in voice agents

* How an end-of-utterance model decides whose turn it is, and the wait time developers can tune

* LiveKit Inference: one key, one contract and no lock-in to a single provider

* Coding agents: "the blind leading the blind", and why he clones the repo into the project

* Real-time avatars, and why people say more to a face than to a waveform

* Personalization as the next horizon for voice agents

* "100,000 minutes yesterday": the growth he sees in healthcare voice agents

* What he would build next: something weird that he feels needs to exist

* Whether voice agents replace call center work or change who does which job

* Hallucinations in regulated industries, and many small agents in place of one giant prompt

* Open source vs. closed source: LiveKit and Pipecat, Retell and Vapi

* The moat when anyone can fork the repository

* Senior and junior engineers: "the toothpaste isn't going back in the tube"

* Why he never wants to do B2B sales again

* Who he talks to on hard days, and why LiveKit feels like a rocket ship

* Why he says judgment is emotional, and the book he recommends

Tom Shapland

* LinkedIn: https://www.linkedin.com/in/tom-shapland-b4494212/

* LiveKit: https://livekit.io/

* LiveKit Agents framework: https://github.com/livekit/agents

* Mentioned in the episode: Homer's Odyssey, the book he recommends; and Never Split the Difference, which Angelina brings up

Timestamps

00:00:00 Intro

00:01:49 Voice AI explained to an 8-year-old

00:04:22 Building a voice agent the naive way

00:07:44 Why your voice agent responds so slowly

00:10:07 Speech-to-speech vs pipeline in production

00:14:32 Why ChatGPT keeps interrupting you

00:19:34 Managing five different API keys

00:22:22 Coding agents build broken voice apps

00:24:42 Why people talk more to fake humans

00:32:30 100,000 voice minutes per day

00:34:08 "Useful" startups fail, "weird" ones win

00:42:07 $1B moat anyone can fork

00:44:18 AI replacing junior engineers?

00:48:47 Why he'll never do sales again

00:56:46 Can a machine have judgment?

01:00:03 Rapid fire + Homer's Odyssey