What I have built

1. NowFounding engineer

   Memfold AI

   Computer use and AI agents, exclusively. Teaching models to operate real software the way a person does, and building the runtime underneath so they can do it unattended.

2. 2025 — 2026Founder

   Midsphere AI

   An agents company. I worked on general agents and the parts they run on: sandboxes, multi-agent networks, and the framework holding them together.

   #22

   Terminal-BenchSeptember 2025

   70.76%

   GAIA, test splitHugging Face leaderboard

   5.5 / 6

   IMO 2025multi-agent net on Gemini 2.5 Pro

3. 2023 — 2024Founder

   Limelight Social

   A gamified video platform for creators. I built the whole thing by hand, without AI: recommendation engine, video processing pipelines, custom models, backend, and the Android and iOS apps. It reached 1,200 users and 2TB of video streaming before I wound it down in October 2024.

   1,200

   users

   2 TB

   video streamed

   1.5 yr

   start to wind-down

4. 2023Solo project

   Hazel.gg

   A gamified social platform: a Flutter client on a Django core, split into separate engines for feeds, progression, and ads, with a topic-classification model layer feeding them.

Open source

Things I have shipped

• mac-cuaPython · Apache-2.0

  A computer-use agent that does not take over your computer. An MCP server driving macOS apps through the accessibility tree in the background — no cursor moved, no focus stolen.

• exoPython · MIT

  An agent framework for distributed systems, with Temporal and E2B support. Memory, context assembly, and token pressure handled by default rather than by configuration.

• optical-read-mcpPython · MIT

  Files handed to a model as densely packed images instead of text. Roughly 7× more code per token on a high-resolution vision model.

Upstream

Contributions

• cua6 PRs1 merged

  Latency work on the macOS driver. Fixed sleeps became configurable pacing, and health checks that ran one after another now run at once.

• open-codex-computer-useSwiftin review

  Background operation on macOS. Drives apps that are covered or on another Space without pulling focus off what you are doing.

• pi_agent_rustRustcredited upstream

  Streamed requests through a LiteLLM proxy were getting a 0% prompt-cache hit rate. Traced it to a missing prompt_cache_key; the fix shipped upstream.