Blog
Notes from the lab.
Engineering notes on computer vision and RAG systems: what we build, how we evaluate it, and where it fails.
Evaluate RAG retrieval before you judge the answers
Most RAG failures start in retrieval, not generation. A practical, step-by-step method for measuring retrieval on its own, then judging answers, then finding out why the system fails.
Building QuarterLens: hybrid retrieval for financial reports
How QuarterLens answers questions about Infosys financial reports by combining keyword and semantic search, reranking candidate passages, and citing the pages behind every answer.
Validate a computer vision project before you build it
Computer vision demos are easy and production systems are hard. A feasibility-first approach: define success, test on real data, find the hard cases, and make a clear go/no-go decision.
Where our football tracker fails, and why
FootballVision detects and tracks players and the ball in match footage. This is an honest look at where it breaks: identity switches, missed balls, and unreliable pitch mapping.