Building in public

Mastermind sessions, in writing.

Every week, Chinat and Xisen sit down to push each other on the companies and ideas they’re building. Each session becomes two short reflections — one from each of us — on what we’re working on, what we’re stuck on, and what we’re committing to.

May 31, 2026Chinat
Talking to Computers in Public
Voice is becoming the dominant way we talk to AI. Cafes are the dominant place we work. Those two things are incompatible right now. The fix is a mask that eats your voice on the way out.
May 31, 2026Chinat
Same Week, Different Bet
Xisen has been running Aicoo for over a month. I shipped Beeper v1 this afternoon. On today's call we discovered we had built versions of the same product from opposite ends of the stack. When two people you trust converge on the same problem, the wave is here.
May 31, 2026Xisen
A Contact Book for Agents
Aicoo is a contact book for agents, not a chat app. Chinat shipped a parallel bet on top of iMessage. The disagreement is where the policy layer lives. The bet is that the agent, not the chat, is the actual product.
May 24, 2026Chinat
Buckets for Rainwater
Hyperscalers won by building the biggest bucket. The next wave wins by placing the bucket where the snowmelt is — asymmetric data, fresh at the source.
May 22, 2026Chinat
Mics in Every Room
Revenue is downstream. The real north star is whether you're collecting data nobody else has.
May 10, 2026Chinat
Public Likes, Private Bookmarks
Demo Day Connect ran for six hours, hit a thousand users, and surfaced an asymmetry between public likes and private bookmarks that no async product can show you cleanly. Live events are an undervalued software leverage point.
May 10, 2026Xisen
Agent Said No, Then Said Yes
Frontier models refuse privacy violations on the first attempt and cave at 2.5x the rate by the fifth. Refusal is a probability floor, not a wall, which means prompt-based policy is structurally insufficient.
May 4, 2026Chinat
The three-hour rule
Cover your week with three hours of paid work, then spend the other one hundred and sixty-five compounding.
April 20, 2026Chinat
Two Hackathons in Two Days
Running back-to-back hackathons at Cornell and Hopkins revealed the difference between a show that needs a showrunner and a machine that needs an operator.
April 20, 2026Xisen
Three Thousand Adversarial Tests
Running 3,000 agent-to-agent security benchmarks revealed the contact security paradox. The same paradox applies to founders: the more consequential your decisions, the more you become the villain in someone's story.
April 12, 2026Chinat
Everything is a CLI now
The shape of every partnership we're signing this quarter is a skill install, not a contract. That's a bigger change than it sounds.
April 12, 2026Xisen
The word we keep coming back to is "networking"
Pulse opened its API this week. The work from here is figuring out what breaks when two agents that don't know each other try to cooperate.
March 14, 2026Chinat
The pipeline is starting to be legible
Demo Day for intake, the Hong Kong flagship for distribution, MentorMates underneath. I'm finally starting to see the shape.
March 14, 2026Xisen
What to benchmark when your users are agents
Pulse security stopped being a concept and started being a benchmark. That changes what the product has to prove.
March 9, 2026Chinat
The flagship is the hackathon, not the product launch
Hong Kong is the operating thread this quarter. Everything else feeds into it.
March 9, 2026Xisen
Why I stopped building an outreach tool
Outreach automation is the wrong shape of problem. The real work is agent-to-agent memory and access.
February 28, 2026Chinat
From a YC hackathon floor to a Hong Kong flagship
I'm on the floor of a YC hackathon this weekend. The real point isn't the prize, it's what the weekend tells me about Hong Kong in July.
February 28, 2026Xisen
Narrowing Pulse to the founder use case
The Pulse pivot is still narrowing. Founders are the sharpest user, and two specific jobs do most of the work.
February 14, 2026Chinat
The Stanford hackathon changed what I thought I was building
A thousand hackers, three days, and one reframe. MentorMates is a platform question, not a feature question.
February 14, 2026Xisen
Two features, done better than anyone
I was chasing differentiation on surface area. The better bet is depth on one or two capabilities.
February 7, 2026Chinat
Partnerships are distribution, not decorations
We shipped the first real agentic layer in MentorMates this week. The lesson: partnerships are only useful when they convert to reach.
February 7, 2026Xisen
User interviews and API-shaped thinking
Two moves that pulled Pulse forward this week: more user interviews, and starting to think about it as an API instead of an app.
February 3, 2026Chinat
Hackathon week is the real test
Two hackathons in one week at Johns Hopkins and Stanford. Distribution showed up. Team reliability didn't.
February 3, 2026Xisen
Heartbeats, active layers, and better testing discipline
Pulse got a heartbeat and an active layer this week. The bigger win was pulling the team into better testing discipline.
January 24, 2026Chinat
Daydreamers Academy and hackathons are the same thing
Sounds like two different products. They're not. They're both structured learning infrastructure — just at different time scales.
January 24, 2026Xisen
The research direction has to be chosen, not discovered
After too many parallel threads I realized the research question doesn't get picked by the field. It gets picked by me.
January 17, 2026Chinat
MentorMates is the center of gravity
A Q1 planning call forced a real choice. Multiple good options, but only one center of gravity. MentorMates wins.
January 17, 2026Xisen
Writing the paper is the discipline
Drafting a paper is a forcing function. You can't hide scope creep when you have to write it down.