Buddy
Maintenance teams were losing time digging through history and documentation. Buddy answers in plain language, cross-checks its sources and cites every claim. I designed, built and shipped it on my own.
この事例はフランス語と英語のみで公開しています。
The problem
In a maintenance shop, the information almost always exists. It is just scattered: intervention history, technical documentation, internal references, each with its own tool and its own way of searching. Finding the right answer means knowing where to look, and which words to use.
An assistant that makes things up is worse than no assistant. An answer without a source is not much better: nobody signs off on a job because of a paragraph that came from nowhere. So the spec fit on one line: every sentence must be verifiable.
Architecture
Documents are chunked, enriched with metadata and indexed twice: once by meaning, once by exact words. The agent queries that index with its own tools, gets the best passages back after reranking, and only answers with what it can cite.
Three decisions
Exact references get lost in pure semantic search. Combining meaning and keywords, then reranking, fixed most of the misses.
Instead of one giant prompt, the agent picks its tools: search, cross-check, summarize. Every step is visible and can be tested on its own.
Over 2,600 tests and 99% coverage gate every release. Evals put numbers on quality, cost and latency before a model or prompt change.
What it looks like
What I measure
Two nets, two jobs. Tests check the wiring: over 2,600 of them, with 99% coverage, and nothing ships while they are red. Evals measure behaviour: retrieval is scored stage by stage on real annotated questions, answers by deterministic checks and a judge with yes-or-no criteria. They don’t vote for me: they give me a table, and I decide.
The set grows with usage: when someone flags a bad answer, it becomes a new test case. It is the least visible part of the project, and the one that saved me the most time.
What I took away
The model matters less than people think. What makes the difference is the quality of retrieval and the discipline of evaluation. And trust comes from the sources on screen, not from how confident the answer sounds.