Projects

Systems I built and what they proved. Most are in bloom, one is still growing. Every number names its source.

shotgrep

Video retrieval · in bloom · Sep 2026

Video search that finds the moment, not the file: fused visual and transcript retrieval over a stage-cached ingest.

habitatFFmpeg · faster-whisper · SigLIP · LanceDB · ONNX · MCP

Recall@5
0.688
MRR
0.495
latency
189 ms p50 / 234 ms p95

measured · frozen eval set, 64 queries over 3,797 moments, Sep 2026

the problem

Searching video usually means searching titles and descriptions. The moments inside stay invisible until someone scrubs through them.

the approach

Stage-cached ingest with FFmpeg, faster-whisper, and SigLIP; fused visual and transcript retrieval over LanceDB; served over REST and MCP tools.

what it proved

Measured, not claimed: 0.688 Recall@5 against a 0.578 keyword baseline, on the same 64 frozen queries.

Contract Change-Impact Intelligence

Retrieval · in bloom · Aug 2026 – Sep 2026

A multi-tenant LLM platform for agreements and amendments: extract the obligations, explain the diff, map each change to the clauses it impacts.

habitatPython · FastAPI · Postgres + pgvector · Postgres FTS · Structured outputs

citation validity
1.0
extraction precision
0.90
recall@10
0.96

measured · project evaluation runs, 2026

the problem

Amendments rewrite agreements a clause at a time. A reviewer does not need a list of what changed; they need to know which clauses each change touches, and they need to be able to check the answer.

the approach

Citation-gated extraction over hybrid retrieval. Every obligation cites the clause it came from, keyword and vector search run together, and per-tenant spend limits keep cost visible call by call.

what it proved

A gate beats a score. Citation validity of 1.0 means the system would rather return nothing than return an answer it cannot support.