I started Retridge because too many AI projects were being judged on demos.

The pattern was always the same. A capable engineering team ships an assistant over internal documents. It performs beautifully in the review meeting. Six weeks later, the people it was built for have gone back to searching the intranet, and nobody can say precisely why.

Not because the team was weak — because nobody had measured anything. There was no query set, no notion of a correct answer, no record of where in the pipeline the answer had gone wrong. Every conversation about the system was therefore a conversation about opinions.

My background is engineering: building systems, instrumenting them, and being on the hook when they misbehave in production. I took MIT xPRO's program in RAG and Context Engineering to formalize that work specifically around retrieval-augmented systems, and built and defended a full production-grade RAG system end to end as the capstone. That is the discipline Retridge runs on.

So every engagement starts the same way: real user queries, scored per stage, with failures assigned a root cause. It is slower than an opinion and considerably cheaper than a rebuild. It also produces something you keep — the evaluation set stays with your team, and can be re-run after every model change.

The consultancy is deliberately narrow. Retrieval, evaluation, security, and production reliability for AI systems grounded in private data. If your problem is somewhere else, I will say so on the first call.