XVANT
Multi-agent orchestration harness
Runs Codex, Claude Code and OpenCode agents in parallel, each on its own isolated Git worktree, with a durable runtime that survives crashes and hand-offs.
- 1,000+ tests
- ~38k lines
- TypeScript
Azrkiel · Mechanical Engineering @ UVA · Agent Builder
I build multi-agent systems that design real, physical things — text-to-CAD assemblies, verified glider wings, benchmarks for AI-generated geometry — and the harnesses that keep those agents honest.
Summon via terminal
$ echo "hi kevin" | mail fac3cq@virginia.edu
What the agent remembers about its operator.
Hermes carried messages between worlds. I carry them between geometry and language — turning what someone means into something that can be machined, flown, or shipped.
I'm a first-year Mechanical Engineering student at the University of Virginia. My research sits where large language models meet physical design: I score AI-generated CAD for BenchCAD, generate and verify glider parts for an early-stage generative-design startup, and build Mirum, a multi-agent pipeline that turns a sentence into a parametric assembly.
On weekends I ship. Two hackathon wins this fall, a 38k-line orchestration harness that runs Codex, Claude Code and OpenCode side by side, and a habit of writing far more tests than anyone asked for.
Loaded from ~/.kevin/skills/
Things I've built and delivered.
Multi-agent orchestration harness
Runs Codex, Claude Code and OpenCode agents in parallel, each on its own isolated Git worktree, with a durable runtime that survives crashes and hand-offs.
Think it. Have it. — text to CAD assemblies
A multi-agent Gemini pipeline that turns plain text into parametric B-Rep CAD assemblies. Fine-tuned Qwen2.5-Coder-7B with LoRA on ABC CAD data to lift valid output from 48% to 78%.
An AI cursor companion for Windows
Lives on your screen, points at things, and draws on top of whatever you're doing — an agent with a body instead of a chat box.
Cheapest model that fits, then escalate
Routes each subagent to the smallest model that can do the job, with structured junior-to-senior escalation hand-offs when the task turns out harder than it looked.
Earlier text-to-CAD experiments for aerospace parts, an engineering-verification toolkit, Hermes agent bridges, the BExSTEM site, and a scholarship finder.
Where agents meet atoms.
Undergraduate Researcher · Prof. Cheng Peng, UVA School of Data Science
A multi-university benchmark scoring AI-generated CAD — v1.0 was used in Anthropic's system card. I'm designing assembly-level scoring for hole alignment and part interference, going beyond per-part voxel IoU.
benchcad.com ↗Intern · early-stage startup founded by Prof. Stephen Baek
I generate parametric glider parts for a pipeline that verifies AI-generated designs before STEP export. Built 12 wing variants — polyhedral wings led on stability checks and calculated glide ratio (11.9 vs. 10.8 tapered). That's the wing spinning at the top of this page.
Independent research project
Multi-agent Gemini pipeline that turns text into parametric B-Rep assemblies — passed 78 of 82 test prompts. Fine-tuned Qwen2.5-Coder-7B with LoRA on ABC CAD data, raising valid output from 48% to 78%.
github.com/Azrkiel/Mirum ↗Senior Capstone · Mathematics & Science Academy
Simulated H2/O2 detonations in OpenFOAM. A 4 mm resonator neck cut peak pressure by ~20%.
The road the messenger took.
B.S. Mechanical Engineering, School of Engineering & Applied Science.
Led 8 staff and grew monthly revenue 23%.
Ran a business & STEM startup competition, fiscally sponsored by Hack Club, for 201 participants in 13 countries. Mentored 3 teams to launch — one to a $200k LOI.
bexstem.org ↗Taught 30 elementary and middle schoolers.
4.71 weighted GPA. Senior capstone on rotating detonation engines.
Led 12 volunteers across 172 hours to build 14 oyster castles for the Virginia Aquarium's shoreline restoration.
Ⅵ Summon
Open to research collaborations, internships, and anyone building agents that touch the physical world.