Angela Wu
Creator of Komo. It began as a side project at Stanford and grew into an AI search engine with 3 million users and zero marketing spend, and one of PCMag's best AI search engines.
I also built the first AI agent that turns know-how into running operations: you write an SOP or record your screen, and it becomes a playbook that automates professional work around the clock.
Earlier, I led engineering and then product on Google Search, working where research becomes product, and took visual search from zero to billions of users.
I angel invest in early-stage AI, focusing on vertical AI applications and AI infrastructure.
Teaching AI to Run Operations
Code got automated because it's formal; operations stayed manual because their knowledge is tacit. This project makes the expertise itself the interface: hand the system an SOP or a screen recording, and it builds and runs the operation.
Rethinking Search Authority
Search ranks by popularity, and people read position as truth; AI answers sound authoritative whether or not they've earned it. This project rebuilds the answer around verifiability: polls that show where sources disagree, reference checks that surface the exact passage behind each claim.
Search Without a Search Box
The query box makes curiosity pay a composition tax. This project makes selection the search gesture: point at any phrase to learn more, quote it, get direct links, or hear what actual people say.
The Mars sample-return mission would collect sealed rock cores from Jezero Crater and launch them to an orbiter for the trip home. Engineers still debate the cheapest way to lift samples off Mars, and researchers hope the cores contain traces of ancient microbial life.
Learning by Building the Map
Understanding is structure, and search sessions end right where structure should begin. One click turns your own thread of questions into a mind map: the shape of the field, traced by your own curiosity.
Can AI Conduct Research?
Models hallucinate most on the questions no careful person would answer in one breath. This project gives the search engine a researcher's method: decompose the question into an editable plan, retrieve evidence per sub-question, synthesize a cited report.
Training LLM with Synthetic Data
RLHF needs crowds of human annotators, which prices out everyone but the labs. This experiment replaced every human in the loop with a stronger model: ChatGPT writes candidate answers, ranks them, and the distilled 131K preference pairs train a 6B student to answer with citations. We got there weeks before Alpaca made the recipe famous.
Search Is Not One Task
Lookup, research, navigation, and opinion-seeking spent twenty years sharing one interface. Komo's founding thesis was to give each journey its own: four modes, four different response shapes.