CS at the University of Washington ('29). I like working close to the hardware: kernels, GPUs, emulators, and robots.

advaithvecham.fly.dev · cv · LinkedIn · advaiv2@cs.washington.edu

• ARC Prize 2026 / ARC-AGI-3 (leaderboard · kaggle: mellamoa). ARC-AGI-3 tests agents on interactive games they've never seen, with no instructions. I'm currently 26th of 3,445 teams (as of Sep 28, 2026). I'm building an agent harness for it. I can't share the notebook or code while the competition is still running.
• Linux kernel, Rust for Linux (fork · regulator.rs · rust-for-linux list). I'm adding safe Regulator methods for current limit, mode, and load on top of the C regulator_* API. The patches go through LKML review.

• SemiLLM (site). Runs LLMs that don't fit in VRAM by streaming quantized layers from CPU to GPU. Each layer's precision is picked on the fly. C++ / CUDA
• weight-switching CoT. When a model switches language mid chain-of-thought, it hot-swaps in a LoRA expert for that language while keeping one shared KV cache. PyTorch
• chip-8 / superchip emulator. A cycle-accurate emulator with a CLI debugger. It passes the Corax+ test ROM. Rust
• Simbiote (event · post). Takes an iPhone LiDAR scan to an OpenUSD sim, trains a robot there with RL, then lets you teleoperate it. Runs fully offline. 1st place, Dell x NVIDIA Hackathon Seattle
• STRUCT (event). Goes from a text prompt to a robot: CAD, PCB, teleop data, and RL, all driven by a 3-bit local model on one DGX Spark. 2nd, NVIDIA Dev Champions @ Spark Hack Seattle
• Insuo (private). A plan-first browser agent with self-healing retries, packaged as a Tauri desktop app. I pitched it at YC Startup School.

• YC Startup School 2026: selected, 5,000 of 35k+ applicants
• IMC Prosperity 4: top 10%
• USACO Gold (solutions)
• AIME qualifier, twice

Rust · C/C++ · CUDA · Python · TypeScript · SQL · Swift
PyTorch · TensorRT-LLM · LangGraph · MuJoCo · Isaac Sim · Kaolin · OpenUSD · Playwright · SDL2 · WebKit
Linux · Docker · AWS · DGX Spark · Git