Hybrid Search Q&A

BM25 + embeddings, fused and re-ranked before it ever reaches the model. Took three rewrites of the chunker to stop losing table context.

RAG Eval Harness

The thing I wish I'd built first instead of third. Now nothing touches retrieval without a faithfulness score attached.

Multi-Step Research Agent

Splits a question into subgoals and hands them to smaller agents. Works well until two of them disagree — still tuning the arbiter.

Self-Correcting Coding Agent

Reads its own failing test output and patches narrowly instead of rewriting the file. Cut regression rate more than I expected.

Defect Detection

Detector trained on maybe 400 labeled frames from the line. Active learning did most of the heavy lifting after that.

Interactive Segmentation Tool

Built on SAM — one click, full mask. Our labelers stopped complaining, which is the real benchmark.

Edge Image Classifier

Distilled a ViT down until it fit on-device. Quantization-aware training saved most of the accuracy we'd have lost otherwise.

Single-GPU QLoRA Pipeline

Fine-tunes a 7B model on one consumer GPU, start to finish. Data prep took longer than the training run itself, as usual.