Michael Chen · AI engineer
I build agent runtimes and evaluation harnesses at Epoch AI / Zerg AI, and I do forward-deployed engineering with our customers.
What I build at Epoch AI / Zerg AI, and the firmware tooling I wrote before it. Some customer details are private.
2025 to now
Zerg is a spec-driven framework for autonomous coding agents, and ZTC is its terminal client. I built parallel tool dispatch so agents stop queueing behind each other, worktree-isolated batch execution so concurrent writers do not collide, and the live web interface where people watch a run as it happens.
ZergChat and Zergboard are built on the same platform.
Epoch AI / Zerg AI · ZTC in TypeScript and Rust, Zerg in Python
2025 to now
For a customer’s organization-management and resource-planning platform, I built continual-validation harnesses that score agent output against deterministic checks of state, tools, artifacts, browser behavior and visual output.
The checks catch regressions before changes are accepted.
Epoch AI / Zerg AI · customer names withheld
2023 to 2025
On a $5M+ DOD-funded fuel-cell program, I built the test-stand apps in C# and Avalonia, wrote a hardware-in-the-loop simulator so firmware could be exercised without a rig, and ported PIC32MZ firmware onto Harmony v3. I also handled migrations and used KiCad and ngspice for circuit work.
50+ modules ported, 2,500+ lines of C tests, and 5 test stands kept running with zero downtime across 18 months.
C#, .NET, Avalonia, Python, PyQt, C, KiCad, ngspice · defense program
A fix I found and sent upstream, merged on 5 August 2026.
I traced five differences between the MLX speech decoders and their PyTorch reference, layer by layer. Every Kokoro render had been 2.5 dB quiet, and the pitch path was one frame out of alignment. The fixes merged upstream.
On a 55-utterance check, MCD dropped from 7.29 to 4.09 dB.
2025 to now
AI engineer, Epoch AI / Zerg AI Agent runtimes, tool dispatch and evaluation harnesses, plus forward-deployed work with customers.
2023 to 2025
MOBIVOLT Test-stand software in C# and Avalonia, hardware-in-the-loop simulation, firmware migration, and KiCad and ngspice work for a DOD-funded fuel-cell program.
2023 to 2025
Founder and president, AI at UCR UC Riverside’s first official AI organization. 50+ members and 10+ technical workshops.
June 2026
B.S. Computer Science, UC Riverside GPA 3.66, with a minor in Strategy & Entrepreneurship.
AMD University Program Award
AMD Pervasive AI Developer Contest, for PHiLIP, a human-in-the-loop image-generation prototype I worked on with a team, built on AMD ROCm.
Cutie Hack 2022
First place in the sustainability track.
Fantasy Sports ·
Week 4 was a 195.18-point win. In Week 5, I kept most of the same starters and lost 133.34–170.60, dropping to 3–2.
Tech & Building ·
I just think the useful interface and the size of the system are separate questions.
Tech & Building ·
A local speech model can ask a stronger model for help and speak its reply while text is still arriving. This is what worked, and where it still falls short.
Away from work, I build apps for friends and personal tools, and I write.
Projects & Fun
Writing About me