Your team takes over, or the engagement failed.

The real problem

A local AI stack delivered without transfer is a liability. It runs while nothing moves, then a model ships, a dependency breaks, a service fails to restart — and nobody on the team knows where to start.

The risk is compounded by the domain: LLMOps skills are recent, poorly documented, and most of the useful knowledge consists of traps no official documentation mentions.

How I solve it

What I actually transfer, in order of importance:

1. Measurement discipline — not commands. Knowing that ps rss lies, knowing to verify a cosine before migrating, knowing to quote a cold latency. That is what travels from one project to the next.
2. The gates — an unmeasured system degrades silently. The team leaves with thresholds that fail red, not with a dashboard someone glances at in a meeting.
3. The domain traps, documented and dated: the ones I paid for, so they get paid for once.
4. Architecture documentation — decisions and abandonments. A rejected option, traced, is worth more than a chosen one with no rationale.

A trap learned in the field

What it delivers, measured

• Bilingual FR/EN architecture documentation, 16 pages, diagrams included
• Traced decision log — abandonments recorded alongside accepted choices
• CI gates delivered ready to wire in, with their baselines
• Recorded handover session, reusable internally

Where it runs in production

• Klody Core — Control plane
• Library Brain — 100% local RAG
• KlodyAI — Coding agent · ReAct