Lead Machine Learning Engineer | Victoria, BC
A first pass at Find-and-Interpret: a sub-1B model fine tuned to locate query-relevant regions in a document page and then filter them for relevance, so a downstream VLM only sees the crops that matter. I walk through the idea, the training setup, and a handful of eyeballed examples showing where it works and where it still breaks.
Text RAG is manageable. Document RAG means bolting on OCR and layout models that work well but add real overhead. Visual RAG skips the extraction step entirely, but trades it for a cost and precision problem of its own.
A chronicle of a reinforcement learning project, from initial struggles with exploding losses to a successful pixel-based agent for LunarLander.
A very approachable jumping off point for video captioning. If you're GPU-poor (<24GB vram) this is for you.
A small contribution to the community. Adds caption-like variety samples to SSV2 dataset.
Some thoughts about the potential power of Meta's V-Jepa 2.