Writing

The full technical write-ups behind the projects — what got measured, what broke, and what the numbers actually mean.

• August 2026 · SEC-RAG-Eval

  SEC-RAG-Eval: measuring RAG where it fails, not where it shines

  The recall number is the least interesting output. What the eval harness caught — a metric that reads 8× different depending on how you score it, and a validation that turned out to be an artifact — is the point.

  Read

• July 2026 · hoopvec

  hoopvec: what CV tracking costs downstream, measured

  Reconstructing player tracking from broadcast video is the easy part. The research question is how much play retrieval degrades on reconstructed tracks versus ground truth — and which perception error actually does the damage.

  Read