SERVICE · 05
Audit & rewrite
We look inside the AI you already run, find where it fails or wastes money, and give you a clear order of what to fix.
You have an AI feature in production: rising costs, answers that don't convince, users who avoid it. With access to the system we measure what really happens, and reach one decision per component: keep, fix, rebuild or remove.
FROM EVIDENCE TO DECISION
Example extract: components and numbers are anonymised and illustrative. The same data follows you down the page.
Example · anonymised data
- —embedding-pipeline—
accuracy 71%, below baseline · loss 2.3 on the test set
wrong answers on ~3 questions out of 10 · effort: medium
REBUILDNew embedding model and document re-indexing - —cost-attribution—
$4.3k/mo unattributed · 60% of the AI budget
no way to know which feature costs what · effort: low
ADDCost tracking per call and per feature - —prompt-routing—
23% of calls go to the premium model · ~$1.2k/mo recoverable
cost without quality gain on simple cases · effort: low
FIXRoute simple cases to a cheaper model - —retrieval-eval—
recall@5 = 0.91 · above target
works as intended · effort: —
KEEPNo change; stays as the test reference
WHAT YOU GET
Not “a big report”: four things you can act on, all in a written report plus a walkthrough session with your team.
- 01
The evidence
For every component: what we measured, how, and what came out.
In the example: 71% accuracy and $4.3k/mo nobody can attribute.
- 02
The priorities
An order by impact and effort, so you know where to start.
In the example: cost attribution first. Low effort, and it unlocks everything else.
- 03
The rewrite plan
What to keep, fix, rebuild or remove, in which sequence, with estimated effort.
In the example: rebuild the embedding pipeline, keep retrieval as the reference.
- 04
Costs and benefits
Where an intervention saves money or improves quality, with the estimate and the assumptions behind it.
In the example: ~$1.2k/mo from routing, with the quality risk made explicit.
HOW WE WORK
Indicative duration for a mid-sized system; we confirm it once we've seen the scope.
- 01days 1–2
Setup & access
Access to the systems, map of the components.
- 02days 3–8
Analysis & evidence
Tests, metrics, in-depth review of every component.
- 03days 9–10
Report & walkthrough
Written report and a session with your team.
WHO IT'S FOR
You have an AI feature in production, costs are rising or results don't convince, and you want to know what to rebuild before putting more money into it.
WHO IT'S NOT FOR
You want to build AI from scratch (→ Operational product builds). You want a theoretical opinion without access to the system: without data there's no audit.
- Indicative duration: 10 working days
- Price agreed before we start
Timelines and terms are indicative. We put them in writing after the first conversation, based on scope.
NEXT STEP
Bring us the system to audit →Tell us what it does, what it costs and what worries you. We reply with what we'd need access to.
If the audit concludes it's better to start over, the next step is a → Operational product build