Process
How an engagement runs
Four phases, each with a defined output. You always know what you are getting, what it costs, and what has to be true before we move on.
AI projects fail in the ambiguity, not in the code. This process exists to remove ambiguity early — before anyone has written a prompt that will not survive contact with production.
Every phase ends with a written artefact you keep: a plan, a working platform, a shipped feature, an operational runbook.
1. Audit — 1 to 2 weeks
I read the code, sample the data, and talk to the engineers who will maintain the result. The output is a written plan: what to build, what to explicitly not build, estimated cost and timeline, and the risks I would not accept.
- System and data-flow review
- Constraint and compliance mapping
- Cost and latency model
- Ranked recommendations
2. Foundation — 2 to 4 weeks
The unglamorous layer: gateway, retrieval, evaluation harness, deployment. Small, testable, reviewed by your team as it lands. This is the phase that decides whether the next eight weeks are calm.
- Platform and gateway in place
- Retrieval pipeline with citations
- Eval suite running in CI
- Deployment and environments
3. Ship — 4 to 8 weeks
One user-facing capability, delivered end to end against a baseline we agreed in week one. Demo-able every Friday, with the measurement already wired in.
- Working feature in production
- Baseline versus result, measured
- Documentation for your team
- Rollout and rollback plan
4. Operate — ongoing
Optional, and where the compounding happens. Monitoring, prompt and retrieval tuning, model migrations, incident response and a monthly written report on quality, cost and usage.
- Monthly quality and cost report
- Model migration when needed
- Incident response
- Roadmap for the next quarter
Principles
Written before built
If it is not in the plan, it is not in scope. Both sides are protected by this, including me.
Measure against a baseline
Agreed in week one, before enthusiasm has a chance to distort the target.
Boring infrastructure
Novelty goes into the product. The plumbing should be something you can debug at 3am.
You own it at the end
Documentation, runbooks and pairing. If I am still indispensable, I have not finished.
Frequently asked questions
Can we start small?
Yes, and most projects should. The audit phase is deliberately cheap and standalone — you can stop there with a plan and implement it yourself, or bring in your own team. No obligation after the first invoice.
How do you handle a fixed deadline?
Tell me in the audit phase and it becomes a planning constraint, not a surprise. I will say plainly if the deadline does not fit the scope, and propose the smallest version that is genuinely worth shipping.
Do you work in our existing repositories?
Yes. Working inside your repository, your CI and your review process is the default — it is how your team keeps ownership, and it makes handover trivial.