# Framewise

- **Event:** [Nebius.Build SF](https://cerebralvalley.ai/e/nebius-build-sf)
- **When:** Sun, Mar 15 at 9:00 AM – 10:00 PM (PDT)
- **Where:** Shack15, San Francisco, CA
- **Team:** [Aishwarya Gune](https://cerebralvalley.ai/u/aishwaryagune)
- **GitHub:** https://github.com/agg111/framewise
- **Demo video:** https://youtu.be/XETQ567842Q
- **Gallery:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery/20

Framewise cuts robot perception costs by 80-90%+ by never sending the same scene twice. Instead of flooding a VLM with redundant frames, it uses pixel-level and semantic change detection to gate inference - only the changed region gets sent, only when something actually moved. On change frames, a fast draft model (Gemma-27B) answers first; if uncertain, it escalates to Qwen-72B. The result: large-model accuracy where it matters, near-zero cost everywhere else.

It implements RAG, reranking, Tavily for more context for robot, streaming frames and measured speculative post draft routing based on certainty (confidence driven).

---

Markdown version of https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery/20. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt