Lab BM25, live

Tune the ranker behind “Ask this site”.

BM25 scores a passage by how often each query word appears in it, how rare that word is across all passages, and how long the passage is. Two knobs control it: k1 caps how much repeating a word helps, and b controls how hard long passages are penalized. Move them and watch the ranking reorder.

Why #1 wins

safety: idf 1.86 × tf 2 → 2.70
classifier: idf 1.86 × tf 2 → 2.70
phrase bonus → 1.50
length 18 words (average 20)

medicationsafetyclassifier
  1. 1
    What I do6.89

    Health AI safety. Classifiers and evaluations that decide when a medical question or answer needs extra care. Keywords: Safety classifiers, Evaluation, Clinical grounding.

  2. 2
    Work at Google4.05

    AI Mode in Google Search (2025, product at Google). Search's conversational mode, built on Gemini. I built the query and response classifiers that keep answers about medication dosage safe.

  3. 3
    About1.57

    Sean Chang is a Research Engineer, Google Research · Health AI. Bringing better health answers to everyone. I build the models, safety systems and infrastructure behind health answers at Google. On the side, I'm building Corneo, a skincare routine app. Based in New York.

  4. 4
    What I do0.00

    Programming languages Sean works in: Python, C++, Java, TypeScript, SQL.

  5. 5
    What I do0.00

    LLM systems and agents. Retrieval that grounds answers in search and literature, and memory that lets agents hand off context. Keywords: RAG, Agent memory, Grounding.

  6. 6
    What I do0.00

    Model training and research. Fine-tuning and controlled experiments, from hypothesis to a measured result. Keywords: Fine-tuning, LoRA / QLoRA, Experiment design.

  7. 7
    What I do0.00

    Cloud infrastructure at scale. Production services on GCP and AWS that stay up when real traffic arrives. Keywords: GCP, AWS, Distributed systems.

  8. 8
    What I do0.00

    Products, end to end. Taking an idea to a shipped app on my own: mobile, web and the backend behind them. Keywords: iOS, Web, Backend.

  9. 9
    Work at Google0.00

    Fitbit Plan for Care (2025, product at Google). A Fitbit Labs experiment that helps people prepare for doctor visits. I built its infrastructure and the conversational memory its agents share.

  10. 10
    Work at Google0.00

    Med-Gemini (2024, research at Google). Gemini models adapted for medicine. I studied answers written for consumers, grounded in fast Google Search and medical literature through a retrieval (RAG) harness.

  11. 11
    Corneo0.00

    Corneo is Sean's side project, an iOS app · skincare routines. Builds a routine from the products you already own, then adjusts it as your skin changes. Recommendations lean on evidence, not hype. Latest: Routine builder from owned products.

  12. 12
    Explainer0.00

    LoRA fine-tuning explained: instead of updating a full d by d weight matrix, low-rank adaptation learns two thin matrices whose product is the update, so only a small share of parameters are trained.

  13. 13
    Agent-native0.00

    Agents and AI assistants can read this site through https://seanchang.me/llms.txt and a read-only MCP server at https://seanchang.me/mcp with tools for profile, skills, Google work, Corneo, updates and site search.

  14. 14
    Now0.00

    Update on 2026-10-09: Rebuilt seanchang.me with GSAP, Lenis and React Bits.

  15. 15
    Now0.00

    Update on 2026-10-06: Corneo: finished the routine builder that starts from products you already own.