Bruce Li · UCLA · Health Intelligence Lab
I build auditable models and the efficient attention systems underneath them — from mass spectra to medical images. Currently at UCLA's Health Intelligence Lab, advised by Yuzhe Yang.
Auditable MS/MS spectrum retrieval.
Represent mass-conserving fragment pairs as sparse evidence tokens before ranking. Closest-mass and overlap-hard negatives, plus evidence-removal tests, ask whether the evidence actually changes the ranking.
Manuscript · under review at AAAI 2027
Pick the key intervals before attention.
A pre-scoring pass that spends a fixed interaction budget only where the signal is. At 131k tokens, ChatGLM perplexity falls from 12.0 to 9.5; this is quality, not a measured speedup.
First author · under review
Dinner from what is already in the fridge.
Inventory, leftovers, dietary constraints, and time treated as one planning problem. Built with Next.js and Supabase; live at howtoeat.food.
Product · 2026
Clinical teaching, inside WeChat.
A Mini Program for intake, a web workspace for review, and a FastAPI retrieval backend — three surfaces serving clinicians, assistants, and students.
WeChat · FastAPI · clinical teaching
Recommendation that knows when to abstain.
Bidirectional interest propagation with exposure budgets; an 11-language audio-visual pipeline that flags conflicting modalities instead of forcing a label.
Recommendation · multimodal · robotics
Co-founder and CTO, building the operating stack.
Data pipelines, workflow automation, and delivery for an AI-assisted housing debt-collection product. Technical direction, hiring, and standards.
Housing · operations · AI
The rest — awards, teaching, earlier roles — is in the archive.
