Research & Retrieval

RETRIEVAL WITH RECEIPTS

The top floor ships AI like infrastructure: deterministic cores, golden Q/A sets, evals gated in CI. If a model can’t be measured, it doesn’t ship.

2PROJECTS ON THIS FLOOR

≥40GOLDEN Q/A PAIRS

0.409MEAN RECALL@5

F32026

CodebaseRAG

Hexagonal RAG over your own codebase, with a deterministic, eval-first core: ingest, query, and CI-gated retrieval evals.

The problem. RAG demos are easy; trustworthy RAG is not. This one pins retrieval quality with a golden Q/A set (≥40 pairs) and fails CI if mean recall@5 drops below its committed baseline.

Highlights

Eval baseline gate in CI: recall@5 0.409 · MRR 0.231 · nDCG@5 0.277 (committed baseline, ADR-gated overrides).
Ports & adapters: pgvector (default) or Qdrant; Ollama or Anthropic embeddings & LLM; Langfuse observability.
Python 3.13 · Pydantic v2 · typer · FastAPI · ruff · black · mypy --strict.

0.409mean recall@5

≥40golden Q/A pairs

strictmypy

Stack: Python 3.13 · FastAPI · pgvector · Qdrant · Ollama · Langfuse · mypy strict

CASE FILE →View on GitHub ↗

F32026

Interview Simulator

Voice interview simulator: a local LLM plays the interviewer and Whisper handles transcription, so you can rehearse technical interviews fully offline.

The problem. Interview practice is awkward with an audience. This one runs fully offline: voice questions, spoken answers and transcription with zero cloud dependency — your awkward answers never leave the premises.

Highlights

Runs fully offline: voice questions, spoken answers and transcription with zero cloud dependency.

0cloud calls

100%offline

Stack: Whisper · Local LLM

CASE FILE →View on GitHub ↗

Back to the building