Sign up to save your podcasts
Or
Share Just Now Possible
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
Just Now Possible
Just Now Possible episodes
- Building an AI Voice Agent for 360 Degree Feedback: How Your360 AI Uses State Machines, Evals, and Layered Synthesis
Guests
- Jared Goralnick, Co-Founder and CEO, Your360 AI
- Eda Goksel Jackson, Head of Design, Your360 AI
- Alex Krasun, AI Engineer, Your360 AI
Key Takeaways
- They built the product one call at a time. The self-discovery agent shipped and was validated a month before feedback collection existed, and synthesis came after that.
- A single prompt couldn't hold a structured conversation. State machines keep the agent on track while still letting it go deep when someone wants to.
- Voice AI is not a survey. People answer three questions in their first sentence, so the agent has to track what has already been covered and decide when to push and when to move on.
- If a customer won't let you look at conversations in the early days, that tells you something about the market. Find design partners who will, then build automated flags so humans only step in when something goes wrong.
- Synthesis needs two kinds of evals. Structural passes catch "don't mess up" errors like misattributed feedback. Substantive passes re-run the analysis with few rules to find patterns the first pass missed.
In this episode
- How interview-based 360 feedback differs from annual reviews and survey-based developmental 360s
- The three calls with Coach Tam: self-discovery, feedback collection, and the 360 report debrief
- Why the first debrief call broke a 60-minute platform limit, and why the median debrief is now 44 minutes
- Using state machines to manage transitions between strengths, growth areas, and requested areas
- Evals built from personas based on failed calls, used to regression-test model and provider upgrades
- Automated flagging on sentiment, customer value, and product quality instead of routine transcript review
- Usability testing a voice call: screen recordings plus follow-up interviews that compare perception with what actually happened
- Moving from Retell to LiveKit for full control over speech-to-text, the LLM, and text-to-speech
- Model tradeoffs for voice: personality vs. time to first token, and why reasoning models get in the way
- Optimizing feedback collection calls for completion and low friction, including calls over the phone
- Anonymity rules: corroboration required for critical feedback, groups shown only at three or more providers
- Progressive reveal in the debrief so recipients process one theme at a time
- Delivering feedback through each company's values and leadership principles
- Organizational heat maps for interdependent teams and independent cohorts
- What's next: turning awareness into action between a person's first and second 360
Resources & Links
- Your360 AI
- LiveKit – the open-source voice orchestration layer Your360 AI moved to
- Retell AI – the voice platform they started on
- Hogan Assessments – the interview-based assessment Teresa compares it to
- Deepgram and ElevenLabs – speech-to-text and text-to-speech layers in a cascading voice stack
- Fearless Feedback by Rebecca Glenn – the coach who helped shape the report format
Chapters00:00 Meet the Team
01:35 What 360 Feedback Means04:11 Who Uses Your 36005:33 Origin Story and AI Fit08:25 How the 360 Works13:55 Making Voice Feel Human22:18 Testing and Evals23:43 Privacy and Observability27:28 Building in Stages32:33 Usability Testing Calls34:31 Testing Call Perception35:12 State Machines for Flow36:58 Flexible Transitions Tuning38:27 Owning the Voice Stack40:42 Model Tradeoffs and Costs42:28 Feedback Interviews Constraints44:49 Synthesis Goals and Anonymity49:11 Progressive Report Delivery52:02 Pattern Based Summaries54:25 Multi Pass QA Framework59:46 Org Level Insights Lenses01:02:45 Why Voice Beats Surveys01:05:58 Roadmap Awareness to Action01:07:30 Closing Thanks1 hr 9 min - Generating Opportunity Solution Trees with AI: How Vistaly Rebuilt Its Product Around Interview Synthesis, Evals, and Re
Guests
- Matt O'Connell, Co-Founder and CEO, Vistaly
- CP Dehli, Co-Founder, Vistaly
- Steve Klein, Co-Founder, Vistaly
- Teresa Torres, host of Just Now Possible, building the AI synthesis services for Vistaly V2
Key Takeaways
- The competition isn't good interview synthesis done faster—it's shallow synthesis that never really happened. That changes what "good enough" means.
- Layered analysis matters more than whether a step is a pipeline or an agent. If the interview snapshot is wrong, every layer above it inherits the error.
- Prompt changes run out. Balancing two opposing error modes took an orchestration change—moving the check into an agentic repair loop—not a better prompt.
- Change sets can't be diffed at the end. The same input/output tree pair has multiple valid change sets, and only the semantically meaningful one makes sense to a user.
- Users don't want to collaborate with the AI step by step. They want the answer, then the ability to correct it.
In this episode
- Why Vistaly rebuilt from scratch instead of layering AI features onto V1—and shut off V1 signups to protect the rewrite
- Rebuilding nearly all of V1's functionality in two and a half months, after three years of building it the first time
- The V1 chat interface that walked users through insights one at a time, and why "still faster than before" wasn't fast enough
- How V2 works: upload three interviews, generate an interview snapshot for each, then synthesize a first-draft opportunity solution tree
- The "house of cards" problem: garbage in at the snapshot layer corrupts the tree above it
- Using a code assertion (too many children on a node) as a cheap pre-filter before spending money on an LLM as a judge
- Four new evals and 16 experiment variants to break a seesaw between missing subgroupings and badly framed parents
- Why the fix was a repair loop in the orchestration, not a better prompt—and how that eval became a production guardrail
- Teaching the agent "the rules of the game" so it logs semantic moves (merge, move, reframe) and produces reliable change sets
- Stable IDs, data provenance, and the "basic computer science" Teresa swept under the rug in the prototype
- What beta data revealed about interview quality: sales demos, stakeholder meetings, and LLM-generated fake transcripts uploaded as "interviews"
- Building a 40-point system for generating realistic synthetic transcripts as test data
- Driving tree edits through MCP, including mapping a full Jira epic backlog into the opportunity space
- Adding an agent back to the canvas, and the versioning and rollback problem that comes with it
- Data residency, SOC 2, GDPR, and moving to European Bedrock inference
- Why model upgrades aren't drop-in: prompts are model- and version-specific, and Bedrock constrains tokens independently of the model
Resources & Links
- Vistaly — Opportunity solution tree software with AI-generated interview snapshots and tree synthesis
- Product Talk — Teresa's AI evals guide and her blog posts on building the Vistaly synthesis services
- Model Context Protocol — How Vistaly users drive tree edits from outside the product
- AWS Bedrock — Where Vistaly runs inference, including EU-hosted inference for European customers
- Just Now Possible episode with Banani — the canvas-plus-agent design UX Teresa references
- Just Now Possible episode with Zencity — the layered data pyramid Teresa references
- All Things Product episode on the quality of evidence, with Petra Wille
Chapters00:00 Meet the Founders
01:05 How Vistaly Began03:10 Community First Partnership04:56 Why Build OST Software07:18 OST Explained Fast10:28 Evidence Inside the Tree12:56 AI Changes Discovery16:58 Chat UI Experiment23:00 Should AI Do Synthesis27:10 Rebuilding as V232:08 Turning Off V1 Signups34:43 V2 Workflow and Tensions39:35 Trusting AI Outputs39:56 Layered Analysis Pyramid41:25 Building Better Trees43:34 Bad Inputs and Coaching47:20 Teaching Through Software50:12 Evals and Guardrails55:23 Deterministic Change Sets01:00:54 UX for Tree Updates01:07:12 Chat Agents and Versioning01:11:17 Privacy Compliance Tradeoffs01:15:06 Model Selection Reality01:17:32 What Comes Next01:19:08 Closing and Links1 hr 20 min - Creating Aha! Builder: Concept to Code, No Engineers Required
Guests
- Brian De Haaff, CEO and Co-Founder, Aha!
- Chris Waters, CTO and Co-Founder, Aha!
- Sarah Moisan-Thomas, Senior Product Manager for Aha! Builder, Aha!
Key Takeaways
- Aha! is fully bootstrapped, remote-first, and profitable—built by product people, for product people
- Containerized infrastructure proved too wasteful for AI app building at scale, pushing Aha! to rebuild on a single-instance, multi-tenant architecture
- Deterministic, pre-built components (authentication, SSO, database, email) cut token cost and guarantee reliability that AI-generated code can't
- A multi-phase, multi-agent pipeline—design system, then prototype, then backend—produces better outcomes than a single open-ended chat prompt
- As building gets easier, knowing what to build becomes the product manager's real differentiator
In this episode
- Aha!'s 13-year bootstrapped, remote, no-salespeople business model
- The five-step discovery framework Aha! uses for every new product, including Builder
- Why their first containerized version was too slow and expensive to scale
- Rebuilding for a single-instance, multi-tenant architecture
- Reusing the agent framework and tool set originally built for "Elle"—Aha!'s roadmapping AI assistant
- Deterministic components for authentication, SSO, database access, and email
- The multi-agent, multi-phase build pipeline: design system, then prototype, then full application
- Enterprise governance: authentication policies, deployment guardrails, and PII/PHI considerations
- Why Aha! is optimizing for prototypes, proofs of concept, and internal "meta applications"—not primary line-of-business apps
- What's next: expanded design systems and tighter feedback loops across Aha!'s product suite
Resources & Links
- Aha!
Chapters
00:00 Meet the Aha! Team
00:19 Aha! Bootstrapped Remote Story01:29 What Is Aha Builder02:27 Leaders Stay Hands On04:49 Origin and AI Shift07:44 Why Build in Aha!14:12 Discovery Framework17:50 From Slides to Early Access21:40 Proof of Concept Approach23:36 Under the Hood Architecture27:01 Agent Tools and UX30:59 Deterministic Components33:42 Skills Docs and Iteration35:48 Higher Level Building Blocks36:33 Knowing What To Build38:57 Builder UX For PMs44:10 Enterprise Governance Guardrails47:29 What Builder Is For50:39 PII Security And Compliance53:15 Quality Evals And Pricing55:29 V8 Isolates Architecture58:22 Multi Agent Build Pipeline01:01:14 Roadmap Design And Feedback01:03:01 PMs Engineers And PR Flow01:05:52 Wrap Up And Thanks1 hr 7 min - Building AI for Women's Health: How Hertility Combined Bayesian Diagnosis and Scan Automation
Guests
- Tulsi Patel, Director of Product and Technology, Hertility
- Lorna Brightmore, Head of Data and AI, Hertility
- Jack Pickard, Head of Engineering, Hertility
In this episode
- What makes Hertility's data set unique: seven years of linked symptoms, blood tests, and pelvic scans from over a million women
- How Gyn.AI uses a Bayesian network to give clinicians probability-based diagnoses instead of binary yes/no calls
- Why showing clinicians the reasoning behind a diagnosis—not just the label—builds trust and speeds up triage
- Guarding against automation bias with holdout sets and independent, fresh-eyes review
- Inside the scan automation pipeline: classifying ultrasound images, detecting follicles, and measuring ovarian volume more precisely than manual methods
- Using an agentic loop to check AI-drafted clinical letters against patient data and catch hallucinations before a human sees them
- The infrastructure challenge of securely piping DICOM ultrasound images from third-party scan providers into Hertility's systems
- How Hertility handles PII and PHI: pseudonymization, data minimization, and running models in-house on AWS Bedrock
- Why treating healthcare regulation as a product requirement from day one makes AI products more scalable, not slower
Key Takeaways
- Probabilistic, transparent AI outputs build more clinician trust than binary classifications.
- Guardrails against automation bias are as important as the model itself.
- Data minimization and in-house infrastructure make it possible to build AI responsibly with sensitive health data.
- Treating regulation as a design constraint from day one makes AI products more defensible and scalable, not slower.
Resources & Links
- Hertility — At-home hormone testing and reproductive health diagnostics for women in the UK and Ireland
- AWS Bedrock — The platform Hertility uses to run LLMs in-house under its own governance and regulatory controls
- PyTorch — The foundation for Hertility's in-house image classification and contouring models
Chapters
00:00 Meet the Team
00:13 What Hertility Does01:51 How Customers Access It04:06 A Unique Women’s Health Dataset07:03 Mission and Efficiency with AI10:03 Why Long Assessments Convert13:52 Before AI Workflows16:52 Research Publications and Impact18:48 GynAI Reducing Time to Diagnosis21:21 Triage and Clinician Support24:37 Keeping Patient UX the Same26:12 Bayesian Network and Explainability30:19 Multiple Diagnoses and Probabilities32:37 Probabilistic Diagnosis Shift33:50 Clinician Adoption and Workflow Fit34:58 Communicating Medical Uncertainty36:43 Scan Automation Overview40:30 In House Image Analysis44:25 DICOM Pipeline Engineering47:30 Evals and Automation Bias50:31 LLM Letter Guardrails56:47 PHI Handling and Regulations01:00:43 Infrastructure Choices and Wrap Up1 hr 5 min - From COVID Pivot to AI World Building: How Snapbar Reinvented the Photo Experience
Guests
- Sam Eitzen, Co-founder & CEO, Snapbar
- Joe Eitzen, Co-founder & CPO, Snapbar
- Patrick Ellis, CTO, Snapbar
You'll hear how they:
- Pivoted from physical photo booths to a cross-platform virtual product in spring 2020 using WebRTC—built from first principles out of necessity
- Integrated Stable Diffusion 1.5 as their first generative AI model and ran custom LoRA fine-tunes on H100/H200 GPUs to produce brand-quality outputs nobody else in their space could match
- Evolved from negative prompts to reasoning model long-form prompts, giving brands more precise creative and safety control
- Built a meta-prompting pre-processing pipeline to ensure user likenesses—including non-obvious details like disabilities—are accurately represented in generated images
- Designed an experiential marketing platform that lets brands "world build" at conferences, trade shows, and live events by bringing fans into branded creative worlds
- Added participatory user inputs through Mad Lib-style prompts and prompt injection, turning photo experiences into co-creation moments between brands and their audiences
- Used Claude Code and Codex to build and ship features rapidly as a small bootstrap team, and developed a four-pillar agent orchestration framework: context, tools, verification, and workflows
- Are building customer-facing "vibe coding" using the Claude Agent SDK so brands can configure and create experiences themselves within Snapbar's platform
Resources & Links
- Snapbar — AI photo booth and experiential marketing platform
- Patrick Ellis on YouTube — Patrick's channel on AI engineering and agent orchestration
- Sam Eitzen on LinkedIn — Follow Sam for experiential marketing insights
- Claude Code — The AI coding tool Patrick and Teresa both use heavily
- Stable Diffusion — One of the generative image models in the mix Snapbar evaluates for its AI photo experiences
- Black Forest Labs / Flux — Another model family Snapbar tests and can route activations to, depending on the creative style needed
Chapters
00:00 Meet the SnapBar Team
00:59 From Photo Booth Side Hustle02:31 Pandemic Pivot to Virtual08:06 What AI Photo Booth Does11:03 Why It Works at Events13:15 When AI Entered the Product14:26 Early GenAI Tech Stack17:43 Patrick’s Learning Path20:44 Why They Bet on AI28:25 Building with Necessity34:25 Applied AI Beats Models36:01 Prompting Got Easier37:53 Arcades to Home Shift38:58 Commoditization vs Application Layer40:58 Experiential Platform Overview42:33 World Cup Activation Examples45:34 QR Flow and WebRTC47:51 Capture to Display Pipeline49:24 Dark Factory and Vibe Coding55:38 Brand Safety Prompting01:00:23 Meta Prompting and Representation01:03:22 Experiential Marketing Momentum01:06:12 Agent Orchestration Framework01:10:02 Claude Code and Fable Rant01:12:48 Wrap Up and Where to Find Them1 hr 15 min - Is This Okay? How Override Labs Built a Safety-First AI Consent Coach for Teen Boys
Guests
- Priya Nakra, Founder and Product Lead, Override Labs
- Olivia Rowley, AI Advisor and Board Member, Override Labs
In this episode
- Why Priya left a 10-year tech career to found a nonprofit focused on gender-based violence prevention
- How scraping 2,000 Reddit posts per subreddit validated demand for a consent reflection tool
- Why Override Labs defined a "South star" — the worst-case outcome — and designed the product to avoid it
- How a licensed therapist and positive masculinity coaches shaped the product's tone and eval rubric
- Why the product never gives a "green flag" response — and what it offers instead
- How risk classification runs deterministically before Claude is ever called, then tailors the response by tier
- The three-part response structure grounded in motivational interviewing: validate, reflect, invite reflection
- Why "privacy by design" meant no accounts, no cookies, no cross-session tracking — and how that became a feature
- The challenge of measuring prevention when success means something didn't happen
- How Override Labs is building a product ecosystem: a women's tool, a web game as a top-of-funnel, and a future institutional API layer
Resources & Links
- Override Labs — Building technology to prevent harm from gender-based violence
- The Fund to Prevent Sexual Assault — The philanthropic incubator that funded Override Labs' first cohort
You can find Override Lab's products here:
- Is this okay?
- Game Boi
Chapters
00:00 Meet Priya and Olivia
00:50 Why Override Labs Exists02:11 From Tech to Mission04:51 Is Technology Good07:49 Narrowing to Teen Consent11:07 Researching Teen Scenarios14:33 Clinical Guidance and Tone16:48 Designing for Reflection20:51 Prototype to Product Flow24:46 Risk Tiers and Evals28:10 Defining the Conversation Goal29:35 Motivational Interviewing Framework31:37 Measuring Real World Impact35:39 Privacy First Data Practices40:29 Reaching Teens With Ads41:56 Web Game Funnel Strategy44:30 Funding and Institutional Pathways46:50 UX for Sensitive Moments49:34 Matching the Older Brother Tone52:09 Building Safe AI for Prevention54:35 Closing Thoughts and Thanks56 min - Beyond Black Box Scores: How Musubi Trains Custom AI for Trust and Safety Teams
Guests
- Nikki Marinsek, Data Scientist, Musubi
- Brian McCaffrey, Software Engineer, Musubi
- Dan Means, Machine Learning Engineer, Musubi
In this episode
- Why off-the-shelf moderation scores fail and how custom-trained models fix that
- How Musubi combines traditional ML with LLMs for different moderation tasks
- The discovery that AI can outperform human moderators—and how to communicate that to clients
- Using AI as a judge to referee disagreements between AI and human decisions
- How Musubi onboards new customers with "reverse demos"
- What custom model training actually means: fine-tuning, feature engineering, and reusable deployment pipelines
- The policy optimizer: an agentic flow that helps customers iterate on their LLM moderation policies
- Why pushing eval tools directly to customers is a core product strategy
- How Musubi is building flexible orchestration workflows for non-technical trust and safety teams
Resources & Links
- Musubi — AI-powered trust and safety toolkit for content platforms
- Maven AI Evals Course — The course Teresa took to learn about evals (get 35% off with Teresa's affiliate link)
Chapters
00:00 Meet the Team
01:18 Why Everyone Wears Product02:32 What Musubi Builds04:51 AI for Human Moderation09:59 Adversaries and Asymmetry11:48 Early Days and Low Latency13:35 First Prototype Slice15:33 Traditional ML Meets LLMs19:52 Benchmarking Against Humans23:09 LLM as Judge and Policy Gaps29:53 From Prototype to Platform31:15 Customer Onboarding Reverse Demos36:08 Custom Models Per Customer38:05 Fine Tuning vs Training39:14 Embedding Driven Classification40:04 Cost and Latency Tradeoffs43:21 Productizing Customization49:16 Scaling Prototypes to Production51:58 Golden Sets and Policy Loops56:17 Coaching Customers Safely01:02:06 Gamified Feedback Signals01:06:19 Agentic Toolkit Roadmap01:09:05 Workflow Orchestration Future01:12:06 Wrap Up and Thanks1 hr 13 min - Building Lorikeet: How AI Humility and a Dual-Agent Architecture Are Redefining Customer Support
Guests:
- Jamie Hall, Co-founder & CTO, Lorikeet
- Xharmagne Carandang, Product Engineer, Lorikeet
- Rona Wang, Product Engineer, Lorikeet
In this episode:
- How Lorikeet evolved from failed ops tools to a full AI customer support concierge
- The dual-agent architecture: Concierge for customer tickets, Coach for configuration and ongoing improvement
- Why "AI humility" — defaulting to human handoff when uncertain — is a core design principle
- How Lorikeet integrates with Zendesk and Intercom instead of replacing them
- The UX evolution from workflow builder to conversational interface — and why the blank chat box is still hard
- "Resolution in the loop": how human agents unblock the AI without taking over a ticket
- Why guardrails need to be domain-specific — the cannabis company story
- How customers define their own evals and guardrails through the Coach interface
- Using AI to diagnose failure modes in traces and automatically suggest fixes
- Lorikeet's product engineering culture: every engineer asks weekly what they learned from a customer
Resources & Links:
- Lorikeet — AI customer support concierge for enterprises in regulated industries
- Gradient Labs on Just Now Possible — another AI agent team in regulated financial services
- Neople on Just Now Possible — AI digital coworkers with a similar training-by-conversation approach
- Incident.io on Just Now Possible — AI SRE with multi-agent hypothesis investigation
Chapters
00:00 Meet the Team
01:05 What Lorikeet Builds02:34 Origin Story and Early Missteps06:42 Finding the Real Support Pain07:37 Why AI Fits Support Work11:16 First Prototype and Early Evals14:42 Design Partners and Selling the CLI16:30 Product Mindset and the Real Moat19:47 Rona Joins and Scaling Up21:02 Milestones Voice Actions Escalation23:48 Integrations with Zendesk Intercom25:59 How the Agent Works Today28:30 Coach Agent and Configuration UX32:58 SOPs to Test Cases34:35 Refund Flow Setup36:12 Coach Conversational UI38:12 Hybrid UX Guidance40:46 Resolution in Loop43:17 Collaboration Middle Ground49:40 Process Maturity Limits53:30 Confidence and Guardrails55:59 Customer Defined Guardrails01:01:14 Trace Diagnosis Agent01:03:14 Product Engineers Culture01:07:46 Closing Thoughts1 hr 9 min - Building Rhea's Factory: How AI-Designed Enzymes Could Finally Solve Plastic Recycling
Guests
- Arzu Sandıkçı, Co-founder & CEO, Rhea's Factory
- Mert Topcu, Co-founder, Rhea's Factory
In this episode:
- Why only 10% of plastic gets recycled—and why mechanical and chemical methods hit a ceiling
- How enzymatic recycling breaks plastic all the way back to its original monomers, unlike traditional methods that just shorten polymer chains
- Why enzymes are selective: they can target specific plastic types even in mixed waste streams
- The discovery of a plastic-eating bacteria in Japan that opened the door to enzymatic recycling
- How AlphaFold and the Nobel Prize in Chemistry transformed what's possible in enzyme engineering
- How Rhea's Factory uses protein language models (PLMs) and multi-step AI pipelines to design novel enzymes computationally
- The evolution from a human-orchestrated pipeline to an agentic AI scientist
- How guardrails at each pipeline step keep the AI pointed in the right direction without limiting exploration
- Why wet lab data—even just hundreds of proprietary data points—can be enough to train a powerful domain-specific prediction model
- Why Mert sometimes wants the model to hallucinate (and how high temperature settings help explore the full enzyme design space)
- The business constraint: enzymatic recycling must compete economically with cheap, oil-based plastic production
- What's next: a process agent, a 5,000-ton demo plant in California, and enzymes for new plastic types
Resources & Links
- Rhea's Factory — Enzymatic plastic recycling technology
- AlphaFold — DeepMind's AI system for protein structure prediction (inspiration for the Nobel Prize in Chemistry)
- Maven AI Evals Course — The course Teresa took to learn about evals (35% off with Teresa's affiliate link)
Chapters
00:00 Meet the Founders
01:50 Why Plastic Circularity03:19 Mechanical vs True Recycling04:52 Biology as the New Tool07:20 Necklace and Pearls Analogy13:22 Low Energy Reactor Process17:33 Origin Story and PET Enzyme22:52 Protein Folding and AlphaFold28:32 AI Designed Enzymes34:28 Protein Language Models Stack37:14 Multi Step Protein Generation39:00 Building on Foundation Models40:50 Lab First Success Metrics43:10 From Human to Agentic Orchestration43:59 Problem Statements as Inputs46:18 Guardrails at Every Stage47:48 Prediction Models and Data Limits50:03 Industrial Reality and Cost52:30 Agentic Parallels and Orchestrators57:45 Impact on Timelines and Diversity01:03:23 When Hallucination Helps01:04:09 Scaling Up and Process Agents01:06:56 Enzyme Blends for Mixed Plastics01:07:49 Why Clamshells Aren't Recyclable01:09:34 Closing Thoughts and Thanks1 hr 11 min - Building AI Employees for Hospitality: How AITropos Takes Orders Where Customers Already Are
Guests
- Santi Marchiori, CEO, AITropos
- Juan Haedo, CTO, AITropos
You'll hear how they
- Spent two years exploring hundreds of startup ideas before finding the specific niche of AI-powered order taking in hospitality
- Went through three product iterations — hardware for waiters, a waiter app, and finally a customer-facing WhatsApp agent — before landing on the right form factor
- Identified order item identification accuracy as their single most important KPI
- Chose a tools-based agent architecture over MCP or pipelines to hit real-time response speed requirements
- Built a parallelized pipeline that searches for multiple products simultaneously and pre-fetches product context before the agent even calls a tool
- Use smaller, fast sub-agents to build an "immediate system prompt" that injects relevant data into each turn without extra tool calls
- Test with thousands of agent-simulated customer conversations run overnight before deploying to new venues
- Reduced new customer onboarding from three months to a few weeks — and continue to shrink it as they build domain templates
Resources & Links
- AITropos
Chapters:
00:00 Meet the Founders00:59 What Tropos Builds01:51 AI vs Human Touch06:17 Restaurant Use Cases08:16 Why Hospitality10:47 Finding the Wedge16:00 Early Prototypes16:46 Hard Parts of Ordering18:03 Speed and Channels21:15 Iteration and Model Jumps30:50 Customer Order Flow35:48 Menu Discovery Question36:07 Menus Inside WhatsApp36:50 Finding the Chat Entry37:37 Why Text Ordering Wins38:30 Under the Hood Pipeline40:54 Tools Over Workflows45:05 Tooling and Prompt Composer49:29 Preloading Context Fast54:02 Founder Learning Mindset57:21 Evaluating Order Accuracy01:00:03 Testing and Human Takeover01:03:56 Onboarding and Scaling Up01:06:10 Whats Next and Wrap1 hr 8 min
About Just Now Possible
From the publisher's feed
How AI products come to life—straight from the builders themselves. In each episode, we dive deep into how teams spotted a customer problem, experimented with AI, prototyped solutions, and shipped…
real features. We dig into everything from workflows and agents to RAG and evaluation strategies, and explore how their products keep evolving. If you’re building with AI, these are the stories for you.