Axiom processes video content, runs GPU inference to predict neurological brain activation in near real-time, and visualizes it as an interactive 3D brain synchronized with video playback.

Everything was built streaming-first — direct browser-to-Cloudflare R2 uploads via Busboy in Node.js, with no files touching local disk. The AI inference pipeline runs on Modal.com GPU infrastructure using Python, keeping it cheap, scalable, and fast.

Early on I hit low accuracy issues with the TribeV2 model. Instead of fighting it, I rethought the approach — generalizing the outputs into an attention graph that maps where in the video the brain is most and least engaged, using signal smoothing to extract meaningful patterns. On top of that graph, I layered Qwen2.5-VL (vision-language) and DeepSeek V3 (LLM reasoning) to generate timestamp-level content edit suggestions grounded in actual neurological engagement data.

SmartClause automates contract analysis — something that traditionally takes lawyers hours. Upload a PDF, DOCX, or TXT, and it extracts clauses, classifies them by risk severity (low, medium, high), and surfaces what matters.

Built on the MERN stack with Mistral LLM APIs at the core. A big chunk of the work was prompt engineering to get accurate, context-aware clause extraction — not just surface-level keyword matching. The REST API layer handles document ingestion and AI inference workflows, and the React + Tailwind UI makes the results clean and easy to navigate.