Weeks 1–2: defining correct
We build the evaluation set with your domain expert: a hundred or more real cases with agreed right answers. This is the least exciting fortnight and the one that decides the project.
In parallel we establish a baseline — what accuracy a straightforward approach achieves — so every later improvement is measured against something.
Weeks 3–4: pipeline and integration
- The extraction or generation pipeline against real data
- Validation rules from your master data and business logic
- Integration with the destination system, including idempotency
- Confidence scoring and the routing thresholds
By the end of week 4 there is something producing output, measured against the evaluation set, with a number attached.
Weeks 5–6: the interface and the guardrails
This is where a demo becomes a system. The review interface, the failure behaviours, the cost caps, the monitoring, the audit trail and the switch that turns it off.
We demonstrate the review interface to the people who will actually use it in week 5, and change it in week 6 based on what they say.
Weeks 7–8: shadow and launch
- Shadow running: the system processes real work nobody acts on
- Comparison against what humans actually did, case by case
- Threshold tuning based on that comparison
- Launch to two or three users, with the switch in reach
- Daily review for the first week, then weekly
Your time, concentrated
| Week | From you | Roughly |
|---|---|---|
| 1–2 | Domain expert on the evaluation set | 2–3 days |
| 3–4 | Questions answered, access provided | 2 hours |
| 5–6 | Interface review with real users | Half a day |
| 7 | Shadow results reviewed together | Half a day |
| 8 | Launch decision | 1 hour |
Booked in advance, that is manageable. Assumed as ad hoc availability, it is the reason projects slip.