How an engagement runs from week one to handover.
Same rails on every project. No surprises in week six. You always know which phase we're in, what the next deliverable is, and what would cause us to stop the line and re-scope. Process is the boring word. Predictability is what it actually buys you.
Four phases. Each one ends in a deliverable you can read in an hour.
SURVEY
Survey the ground before we move dirt.
Two weeks of paid discovery. We read the code, sit in on standups, interview the people who actually use the system, and put numbers on the things you only knew by feel. The output is a scope document with risks and assumptions named out loud. Yours to keep, whether you hire us or not.
- WK 1 D1–3Kickoff, system access, observation
- WK 1 D4–5Interviews + data sampling
- WK 2 D1–4Drafting + technical risk register
- WK 2 D5Scope walkthrough + fixed proposal
FRAME
Frame the architecture before we ship a line.
One week to lock the architecture, data contracts, model approach, and the evaluation plan. A short document your board can read and your engineers can build from. We disagree out loud, in writing, at this stage. It costs nothing here. It costs everything in week eight.
- D1Architecture brief
- D2Data + model contracts
- D3Evaluation + acceptance criteria
- D4Security + compliance review
- D5Sign-off + sprint zero plan
SHIP
Ship in weekly increments, on a real pipeline.
Demos every Friday. Deploys at least every other week. Real users on staging by week four. We treat a scope change as a signal, not a sin. When reality disagrees with the plan, we re-plan. Budget for change is set in phase two and tracked openly so nothing surprises anyone on either side.
- WEEKLYDemo + decision log
- FORTNIGHTProduction deploy
- WK 4Staging open to your team
- WK 6First prod traffic, gated
- WK 8+Full traffic + load tests
HARDEN
Harden, document, hand off, actually leave.
Two to four weeks of paired work with your engineers. Load tests, drift monitors for the ML side, on-call runbooks, incident response drills, and a documented escalation path. We leave when your team signs off, not when the contract ends. Sometimes earlier.
- WK -2Load + chaos testing
- WK -1Runbook + on-call drills
- WK 0Handover sessions, recorded
- +30DStanding office hours
- +90DPost-mortem + invitation back
Six rules we ship by, written down on purpose.
Boring on purpose.
We pick the smallest tool that wins the benchmark, the simplest pipeline that meets the SLO, the smallest model that beats the baseline. Novelty for its own sake is a tax our clients shouldn't pay.
The dashboard is the proof.
If we built it, you can see it run. Throughput, latency, error rate, drift, cost per request. No black boxes. No "trust us" reports.
Models are software.
Version-controlled, reviewed, evaluated against a held-out set, deployed through CI/CD, rolled back like any other artifact. ML stops being magic the moment it has a test suite.
Disagree early, in writing.
If we think a decision is wrong, we write it down before the build starts. You sign off knowing exactly what we'd have done differently. That note is in your repo a year later, when it matters.
Hand it off, then leave.
The exit is part of the design. Every project ends with paired work, documentation, and a paper trail your engineers can actually use.
Talk to the user.
The warehouse lead. The nurse. The support agent. The person who opens the screen at 7am gets a seat in every scoping conversation. The work is for them, not for us or you.
Ready to survey something?
Two-week paid discovery is $9,200. You leave with a fixed-price scope, an architecture brief, and a risk register. Whether you hire us for the build or not.