The frontier belongs to what compounds

I've joined micro1's research engineering team because the next era of AI will be won by teams that build scalable, stable systems for the long term.

My background spans robotics, machine learning, and mechanical engineering, across ML, software, and data engineering roles. Most recently, I led engineering for multimodal pseudo-anonymization and synthetic data generation, built from zero to one. Over my career, I've arrived at a few convictions.

Rigor is the shortest path to the frontier

The hardest problems in AI infrastructure are research questions, not tickets. Every modality is its own field, deep enough for entire PhDs, and a multimodal system is a portfolio of research problems sharing infrastructure.

Treating it that way starts before the first line of code: a quality bar with acceptance criteria, an evaluation harness built before the system it measures, baselines and ablations instead of intuition, and design reviews that record alternatives and failure modes. Failures are studied as classes, not patched one at a time.

Timelines follow the difficulty of the problem, not the urgency of the quarter. Adding engineers to an unframed problem doesn't make it faster. It makes it more expensive.

Velocity is speed with direction

Engineering work either compounds or gets redone, and which one is decided before the first deadline. Compounding takes a north star, a multi-quarter roadmap, and an architecture designed for the next order of magnitude: versioned data with lineage and evaluations that run on every change.

With them, every delivery leaves reusable assets behind, and the tenth costs a fraction of the first. Without them, every deadline gets a bespoke pipeline deprecated at the next one, and the team mistakes motion for progress. The most consequential decision an engineering leader makes isn't what to build next. It's what the team will still depend on a year from now.

Revenue is a verdict, not a roadmap

Count datasets delivered and you get datasets. Count this quarter's revenue and you get this quarter's revenue, borrowed against every quarter after it. A safeguard is not a feature to drop the moment a buyer stops asking for it.

Candor is infrastructure

The most mature thing an engineer can do is surface a risk early, quantify the tradeoffs, and bring it to whoever owns the decision. Organizations that treat that as a problem of tone get fewer warnings, not fewer problems. Engineering judgment can't be managed from a revenue dashboard.

Talent strategy begins after the offer

Prestigious employers, schools, and experience at scale are the easy part of hiring. The real test is what an organization lets that talent do after the offer is signed. Hire a scientist and give them only tickets, and you've paid for a scientist and received a ticket-closer.

Why I'm betting on micro1

micro1 is building research-grade data infrastructure for frontier AI. The companies that win this market won't be the ones that ship the most. They'll be the ones that build what compounds.

If you build with the same conviction, I'd love to talk.