Jimmy Yao
Co-founder, Voluma Bio
Previously AI/ML at Epic Systems
Co-founder, Voluma Bio
Previously AI/ML at Epic Systems
I’m a software engineer focused on applying AI and LLMs to the real world.
Right now I’m co-founding Voluma Bio, where we’re building agents to take over the computational analysis after a lab experiment: segmenting the scans, placing the landmarks, running the measurements, with a reviewer agent checking each step.
Before that I spent three years at Epic Systems shipping LLM summary workflows for clinical chart review across 100+ hospital systems.
Away from the keyboard: tennis, the great outdoors when I can get out into it, and the firm position that Roronoa Zoro is the GOAT.
Building AI agents that automate the computational analysis after lab experiments.
Shipped LLM summary pipelines for clinical chart review across 100+ hospital systems, summarizing 300 million notes in one month.
Built an AWS-deployed UX interview analysis pipeline (Transcribe, BERT/T5 summarization, sentiment, topic clustering) with a React dashboard for reviewing results.
Owned the health-data backend (Flask, Docker, MongoDB, AWS), built the aiSpine posture-wearable Android app, and ported speech and facial stroke-screening models to a Flutter app with TensorFlow Lite.
BS Computer Science; Minor in Mathematics and Business
Launches and tracks specialist agents across Claude Code, Codex, OpenCode, Cursor, and Pi, with session forking, recursive delegation, and file-based memory for long-running agents.
A package manager for agents, skills, and hooks: resolves a lockfile, compiles harness-native configuration, and never deletes files it did not create.
A Mars package of agents and skills for long-form fiction: planning, drafting, critique, and continuity.
An AI writing platform where each agent thread is a Yjs CRDT peer, editing live documents alongside humans or a review draft the writer applies atomically.
An offline iOS read-aloud app running Phonon TTS on Core ML and Kokoro on ONNX Runtime. Moving Kokoro off Core ML cut its crashes from 5 in 20 calls to 0 in 320, at 3.3× real time on an iPhone 17 Pro.
A phone-first tmux client. A Rust daemon fans tmux control-mode output to web and iOS clients with per-connection backpressure and snapshot reconnect.
Benchmarked BM25, hybrid retrieval, and tool-using Gemma 4 agents on TemporalWiki. Agents reached 95% temporal recall@5 but about 20% exact match: evidence use, not retrieval, was the bottleneck.
Co-first author. Contributed wet-lab experiments, data analysis, and manuscript writing.
Added .epub ingestion to LlamaIndex/LlamaHub using EbookLib and html2text.
I scraped 3 months of AI agent sessions across Claude Code, Codex, and OpenCode. 75% of file reads were redundant.
A proposal: let the harness hold one current copy of each active file, instead of storing every read in the conversation.
198 agent spawns ran overnight and half my Codex plan was gone by morning. The culprit was a parent agent that couldn't just wait.
Jimmy Yao builds AI agents for lab science at Voluma Bio. Previously shipped LLM workflows for clinical chart review at Epic Systems.
haowjy.github.io