How we engage.
Four phases. Six to twelve weeks for most projects. Eval-driven from day one. The team you hand off to is part of the design.
Discovery
What problem, for whom, with what constraints? We read the data, talk to the people who’ll use it, and write down what success means. If we don’t think we can ship it, we say so.
Evaluation design
Before we write retrieval or pipeline code, we build the evaluation set. Real inputs, hand-labeled, versioned. It is the contract for the rest of the project.
Build
Working software shipped behind a flag at the end of week one. Iterations are scored against the evaluation set. We refuse to call something "done" without a regression test for it.
Handoff
Documentation, runbooks, dashboards, on-call rotations, evaluation harness, and a 30-day stabilization window where we’re responsive to incidents. Then we rotate off — your team owns it.
What helps us help you faster
You don’t need a perfect brief. The list below shortens our discovery phase by half — but if you’ve got nothing more than a one-paragraph problem statement, send that.
- The problem in two paragraphs. The user, the friction, what done looks like.
- 20 example inputs. Real ones, not synthetic. With expected outputs if you have them.
- Constraints that aren’t negotiable. Latency budgets, regulatory requirements, on-prem requirements, vendor lock-in concerns.
- The team that will own it. One name is fine. We’ll work with them from week one.
- Small team. One or two engineers, deep ownership. No "delivery managers." No "AI strategists." No layers.
- Your stack. Python, TypeScript, Postgres, and your choice of Claude, OpenAI, or Gemini — or a routed mix. We adapt to the provider that fits your problem; you don’t learn a new platform to keep what we ship running.
- Async by default. Daily written updates. Weekly demo. Slack/Teams during your hours. We’re available, not intrusive.
- Fixed scope, weekly milestones. The first week of the build phase is the most important: by Friday, something is in front of a user.
Send the problem.
Two paragraphs. We’ll respond within one business day with whether we’re a fit and what we’d look at first.