Djaouad Frih.

Engineering graduate from ESI SBA, working on one thing: putting AI into systems that hold up in production. I design, build and deploy the whole stack, from vector retrieval and agent tool-calling through to auth, tenancy and deployment. Two LangGraph agents and two production retrieval pipelines are running on real domains now, all with public repositories.

I hold a line on measurement. Retrieval is scored offline against a goldset rather than assumed. The benchmark runs on the AI Virtual Receptionist knowledge base (dental clinic KB, 15 chunks, 18 queries) with a deterministic offline embedder (concept lens + hashed lexical residual). The most recent run reached 93.8% F1 at threshold 0.35. This measures the retrieval policy (the similarity floor, top-k cut, and vector-vs-keyword choice), not the production embedder (text-embedding-3-small) or end-to-end answer quality. Token usage and latency are recorded per request and served through an authenticated metrics endpoint. Where something is not measured, I say so rather than implying it is.

On tenure: these systems were built between August and October 2026, and the receptionist is the one I have kept running longest. I am early in my career and I would rather say that plainly than dress it up. What I bring is depth on the AI-systems side and the ability to own a build from schema to production without handing it to someone else.

Not public projects, and listed separately for that reason. These have no public repository and some involve outreach tooling, so I am describing the capability rather than inviting you to inspect a repo.

• A B2B sales-intelligence pipeline of 12 MCP tools behind a 9-step orchestrator, with an automated QA gate before anything is sent. Internal tooling, not a public repository, and not listed above for that reason.
• A lead and outreach MCP server of roughly 60 files covering discovery, scoring and sequence management. Internal, and deliberately not public.
• An offline eval harness that scores retrieval against a goldset, with a threshold sweep to pick the operating point.
• Token usage and latency capture persisted per request and served through an authenticated metrics endpoint, so a running system's cost is readable rather than guessed.
• A content pipeline with a human approval gate: strategy, generation, automated quality check, human sign-off, publish, analytics, and the results fed back into the next cycle.