Superhuman charges $30/month for email when Gmail is free. That's always forced them to maintain a different quality bar from most products, and it shapes everything about how they build AI features too.

Loïc Houssier is CTO at Superhuman Mail, and one of the most fun and energized engineering leaders I've gotten to work with. In this conversation, he walks us through what quality really means when you're building a "luxury" software product - and how that mindset applies to AI.

We dig into the high-dimensional challenge of building great AI experiences around email, from auto-drafts to semantic search. We talk about how they approach evals when every user's inbox looks completely different, starting from the hardest queries they can find internally. And we get into how Superhuman is adopting coding agents across their engineering team - including their "quality week" practice and why they removed all procurement blockers for AI tools.

Timestamps: 

0:00 - Intro: What "luxury" means for a software product 

2:12 - Loïc's background and why he joined Superhuman 

6:02 - Game design principles in product development (not gamification) 

10:14 - AI features at Superhuman: triage, search, and auto-drafts 

15:34 - The challenge of building AI features in a high-dimensional space 

20:00 - Building evals from the hardest internal queries (the "wood for my coffee table" example) 

24:20 - Privacy and how they handle eval data 

25:45 - Their mix of models: BERT, fine-tuned, open source, and frontier 

28:19 - What quality means when your competition is free 

30:10 - Quality week: dedicating the first week of every quarter to bugs and AI workflow improvements 

32:44 - How they're adopting coding agents internally 

35:30 - Removing all the blockers for AI tools (e.g. 24-hour security approval, unlimited budgets) 

38:06 - How Loïc ramped up on AI as a leader 

40:15 - Parting advice: choose your vendors wisely, and enjoy this moment as a builder

Links:

• Superhuman: https://superhuman.com
• Find Loïc on LinkedIn: https://www.linkedin.com/in/loichoussier/

On this episode of Deployed we talk with Lijuan Qin, Head of Product for Zoom AI, about how her team is moving beyond AI meeting transcriptions and note-taking to a mission of agents helping from "conversation to completion." That's how Lijuan describes her vision of the future for Zoom AI, where AI doesn't just summarize your meetings, it actually follows through on the work that comes after.

Lijuan has a PhD in AI and spent 20 years at Microsoft working on NLP and video understanding before joining Zoom. She brings a long-arc perspective on what's changed and what hasn't in AI, and shares how her team thinks about building an AI companion that acts more like a team member than a search engine.

Key insights for builders include:

* Why high engagement with an AI product can be a negative signal. i.e. if users keep going back and forth with your AI, the product might be failing them ("you got it wrong! try again").
* How Zoom measures AI quality by *output* value and task completion, instead of usage metrics or individual response accuracy
* Their "AI-first, intent-driven" approach: starting from what the user needs to get done, not which tool to use
* How they personalize AI features in stages: role-based outputs first, then memory, then live conversation context, rather than trying to build something like a full "digital twin" on day one
* A concrete example: Drafting different kickoff documents for each meeting attendee that is personalized for the priorities of their role (PM vs. engineer vs. CTO)
* Why transparent decision frameworks let big organizations experiment fast without approval loops
* How the Zoom AI team balances speed and enterprise trust

Links from our conversation:

* Zoom AI Companion: https://ai.zoom.us
* Find Lijuan on LinkedIn: https://www.linkedin.com/in/lijuanqin/
* Freeplay (that's us): https://freeplay.ai

Kevin Stanton has spent 13 years at Sprout Social, most recently running infrastructure for a platform that processes billions of social posts. When generative AI emerged, their team saw an opportunity to solve one of their hardest problems: helping customers make sense of massive amounts of unstructured social data.

Now Kevin is building Trellis, Sprout's AI agent for social listening and competitive intelligence. In this conversation, he shares what it's looked like to shift an engineering team toward building agents — and the practical lessons they've learned shipping to thousands of customers.

We cover details like why MCP felt more natural than RAG for their architecture, how they use chat as a strategy for seeding eval datasets, when to let agents reason versus when to collapse tools and write deterministic code, and why they pulled evals out of CI/CD after learning the hard way how non-deterministic tests can break things.

Links from our conversation:

• Sprout Social: https://sproutsocial.com
• Sprout Social Insights Blog: https://sproutsocial.com/insights
• Trellis: https://sproutsocial.com/insights/press/sprout-social-unveils-trellis-its-ai-agent-that-turns-social-data-into-instant-enterprise-intelligence/
• Find Kevin on LinkedIn: https://www.linkedin.com/in/kevinstanton/

On this episode of Deployed, we sit down with Shuhao Zhang, co-founder and CPO of TinyFish, on launch day for Mino (mino.ai) - an enterprise web agent platform that handles reasoning, multi-step execution, and parallel browser sessions at scale.

Shuhao shares his vision for the "operational web" where AI agents become the primary operators of the web, unlimited by human constraints. He shares a live demo and customer stories, like working with Google to connect tens of thousands of Japanese hotels to consumers, and with ClassPass to maintain up-to-date pricing across 30,000 wellness studios.

Key technical insights for builders include:
- Why public benchmarks for web agents are saturated today, and you need your own evals
- How they built infrastructure to replay sessions like a "time machine" for data iteration
- Why they fix issues through data improvement rather than custom patches / rule-based systems

Learn more:
- Minnow: mino.ai
- Tiny Fish: tinyfish.ai
- AgentQL: agentql.com
- Find Shuhao on LinkedIn (https://www.linkedin.com/in/shuhao/) and Twitter (https://x.com/shuhao_friday)

On this episode of Deployed, we sit down with Ajit Varma, VP of Product at Mozilla Firefox, and former product leader at Google, Meta, Square, and WhatsApp. 

Ajit brings a unique perspective: while venture-backed AI browsers race to build what increasingly look like walled gardens, Mozilla has spent months quietly shipping privacy-first AI features that put user choice above everything else.

It's a thoughtful conversation with candid insights into why Firefox offers multiple AI chat providers in their sidebar instead of forcing users into a single model (and how this philosophical stance shapes every technical decision they make), the reality that privacy-preserving AI requires completely rethinking architecture, and why their AI features prioritize on-device processing whenever possible. 

We also get into the technical challenges that come with balancing local and cloud AI (users want SOTA quality but also want their data to stay private), the business model constraints of being a nonprofit foundation competing against VC-funded competitors who can burn cash on flashy features, and why Mozilla believes keeping the web open matters more than winning the AI browser wars.

Whether you're interested in privacy and the open web or just trying to understand what it takes to integrate cutting-edge AI into existing products, Ajit's perspective offers practical lessons from someone navigating a challenging balancing act.

On this episode of Deployed, we sit down with Surojit Chatterjee, CEO and founder of Ema, and former VP of Product at Google and Chief Product Officer at Coinbase. Surojit brings a rare perspective: he started building enterprise AI agents in early 2023—well before "agentic AI" became a buzzword- and has spent two years getting them to production quality that companies like Hitachi actually trust. 

Surojit gives us refreshingly candid insights into why Ema calls their products "AI Employees" instead of agents (and how this completely changes their approach to feedback and evaluation), the reality that most enterprise AI projects fail because companies try to automate broken processes instead of redesigning them, and why multi-agent systems are essential for handling real enterprise complexity. We dive into the technical challenges that remain unsolved (complex dynamic planning with long tool chains is still hard), his forward-deployed "agentic clinics" approach to customer success, and why sustainable scale matters more than flashy demos. 

Whether you're building AI products for the enterprise or trying to understand what it really takes to transform critical business processes with AI, Surojit's battle-tested perspective offers practical lessons from someone who's been in the trenches longer than most.

In this episode of Deployed, we sit down with Ryan Carson, Builder in Residence at Amp (Sourcegraph's coding agent) and founder of Treehouse, which taught over a million people to code. Ryan brings a rare dual perspective: he's both building his own company with AI tools and helping create enterprise-grade AI coding infrastructure.

Ryan gives us refreshingly honest insights into the hyper-competitive coding agent landscape, why traditional evals don't work for open-ended coding tasks, and the "dirty secret" that most quality decisions still come down to "dev vibes." We dive into the technical reality of competing with well-funded teams, his systematic framework for building with AI agents (which has 5,000+ GitHub stars), and why success isn't just about model capabilities — it's about solving real developer problems through obsessive attention to user experience.

Whether you're building AI products or trying to understand what it really takes to compete in this space, Ryan's grounded perspective cuts through the hype with practical lessons from the trenches.

Retry

On this episode of Deployed, we talk with Kelly Schaefer, a Product Director at Google Labs and who’s been recognized as one of the Top 100 Women in AI. Kelly has led a portfolio of experimental AI products like NotebookLM and the Jules coding agent. The Google Labs team helps turn DeepMind research into real products that can work at Google scale.

She shares what actually works when shipping AI features, from how her teams use evals to drive product quality, to why speed matters more than perfection. We also talk about how PM and UX roles are evolving in the AI era, and why she hires “dot connectors” who can bridge across domains.

If you’re building AI products, managing product development teams, or just trying to stay ahead of how this field is evolving, this conversation offers a clear look into what actually works inside one of the world’s most influential AI product orgs.

To see more of Kelly and her team’s work, check out https://labs.google/

On this episode of Deployed we talk with Kwindla Hultman-Kramer, co-founder of Daily (daily.co) and creator of Pipecat (pipecat.ai), the most widely used open source framework for voice agent orchestration.

Kwin shares insights from building voice AI infrastructure since before it was cool, including why he thinks we've hit an inflection point now where voice agents are quickly moving from demos to real production deployments with real ROI. 

He breaks down the technical stack that actually works in production in July 2025, explains why most audio-specific evals are still "vibes" (and why that's okay if you get your text evals right!), and shares tactical advice that could save months of trial and error — like why you should use WebRTC instead of WebSockets, and why speech-to-speech models aren't quite ready for production yet.

Whether you're curious about voice AI or already building voice agents, this conversation offers practical guidance from someone who's seen hundreds of teams navigate the journey from prototype to production scale.

If you want to go deeper on this content, check out Kwin's Voice AI & Voice Agents book (https://voiceaiandvoiceagents.com/) and his popular Maven course (https://maven.com/pipecat/voice-ai-and-voice-agents-a-technical-deep-dive).

On this episode we talk with Nathan Sobo, co-founder and CEO of Zed, the high-performance code editor that's reimagining how AI agents can improve developer workflows. 

Nathan shares lessons from building one of the most natural agentic coding experiences available, including why investing up front to automate the feedback loop to improve the quality of AI systems is worth it. He also shares some great product design insights that go beyond the code editor -- including how they were able to weave an AI agent into the UX of their already-collaborative product (just like a human collaborator), and how their "subtle mode" for code completion has helped win over AI-skeptical developers. 

Whether you're building coding tools or another AI product, this conversation is a fun listen for anyone who cares about crafting great products and incorporating AI in smart ways.