Two million users, one agent: the failures we did not test for

Corvid's design team grew twentyfold in five years and the work got better. Anika Sharma opens Summit with the decisions that kept quality alive, and the ones that nearly didn't.

Corvid's research agent passed every evaluation, then met two million real people. Kenji Arai on the failures no test suite caught, and the rollout system built in response.

Halden moved from one large database to several smaller ones without downtime. Lucas Moreau on the three steps he would do differently, and the one he would repeat exactly.

A hands-on workshop in building prototypes from real, messy data, so your designs meet long names, empty states and awkward numbers before your engineers do.

Orbital used to publish a quarterly roadmap and deliver half of it. Grace Liu on replacing it with a short list of commitments the company keeps, and what that did for trust.

Plenty of research is presented and forgotten. Inês Carvalho on how Quill's research team made its findings hard to ignore, starting with a study of why learners quit in week three.

AI tools have changed what one person can build. A founder, an investor, an engineering director and a product leader on what that means for hiring, team shape and ambition.

Three quick talks on small product changes with outsized results: a changed default, a five-person test and a changelog people actually read.

Stackwell serves forty million requests a day on a stack most teams would call dull. Daniel Reis on why that is a deliberate choice, and the three times it nearly wasn't.

In one year Tidewater's collaboration products removed more features than they added. Elena Petrova on how they chose, how customers reacted, and what it did to support and speed.

Bring a feature that uses a language model and leave with an evaluation suite for it: real test cases, graded outputs and a score you can track from one model change to the next.

A model that takes eight seconds to answer is a design problem as much as an engineering one. Omar Saleh on AI interfaces that feel fast on slow phones and slower networks.

Monogram's engineering team grew from twelve people to two hundred in four years. Noah Whitaker on the structural choices that were easy to change, and the few that were not.

Halden's AI assistant drafts replies for support agents at twelve thousand companies. Zara Haddad closes day one with what two years, two model changes and a public apology taught her about trust.

Stackwell cut the median time from sign-up to a working deploy from twenty-six minutes to four. Jonah Adeyemi on the measurement, eleven rewrites and the fixes that were not in the docs.

Lumio serves three million small traders on low-cost phones and patchy networks. Maya Okafor opens day two with what designing under real constraints can teach every product team.

Lumio replaced its dashboard's single-page app with server-rendered pages and small interactive islands. Omar Saleh on the migration, the numbers and the parts that were harder than expected.

Kitebird became the first Nordic ticketing app to pass a full accessibility audit. Hannah Berg on why fixing it late cost four times as much, and how to stop the debt building up again.

A hands-on workshop in adding tracing, structured logs and useful metrics to a small but realistic app, then using them to find three bugs planted in it.

Northbeam's design system serves fourteen product teams with a team of nine. Priya Nair on running it with a roadmap, a support rota and adoption figures leadership actually reads.

Public benchmarks say models are nearly perfect at analysis. A benchmark built from real company data says otherwise. Elif Demir on what it takes to measure models honestly.

AI agents now take actions, not just make suggestions. A product manager, an AI product lead, an ML engineer and a CTO on who owns the outcome when one of them makes a mistake.

Four short talks on small interface details that carried more weight than anyone expected, from a loading state to a single line of microcopy.

Wei Chen has read hundreds of board decks. He explains which product numbers change an investor's mind, which get ignored, and why retention curves come before revenue.

Most developers meet your product through its API reference, not its homepage. Jonah Adeyemi on how Stackwell rebuilt its reference around tasks, examples and errors.

A hands-on workshop that compresses product discovery into two hours: frame the problem, talk to real people, test a rough prototype and decide what to build next.

At Arbor, engineers own outcomes rather than tasks. Tomás Ferreira on what that changes day to day, from talking to restaurant owners to deciding not to build what the ticket asks for.

Halden's European customers asked for their data to stay in-region. Lucas Moreau on swapping one large hosted model for small ones in six regions, and what it did to cost and quality.

Amara Boateng closes Summit 2027 with what building Fernway taught her: the way a company works is something you design, ship and maintain, just like its software.

Orbital's prices used to be set in a spreadsheet once a year. Grace Liu on bringing pricing into the product team, and how a new model changed what customers used and what got built.