Evan Chisholm

Computer Science, University of Waterloo

I like building things from scratch, ideally close to the metal: inference engines, renderers, editors, and the agent systems on top of them.

Work

Software Engineering Intern, PathAISep 2026 to now Ingestion of whole-slide images and HL7 data into AISight. Prototyped in-browser inference for pathology ViTs on custom WebGPU kernels, 2x faster than PyTorch MPS.

Software Engineering Intern, MashJan to Apr 2026 A production multi-agent system over 10+ data sources giving revenue teams cited answers, plus the eval pipeline that tracks its regressions.

Fullstack AI Software Engineer, AdanomadMay to Aug 2025 MCP integration for AI agents across 500+ servers, and an 11x Postgres speedup for an AI CRM.

Projects

STEELC++ LLM inference engine from scratch: GGUF loader, BPE tokenizer, KV cache, hand-written AVX2/FMA kernels and a custom thread pool. 20% faster than llama.cpp on FP32 decode.

VibeMovieTypeScript AI video editor. Natural-language edits become schema-validated mutations of a declarative JSON video, compiled to React and rendered to MP4 with Remotion.

vmC++ A Vim-like modal editor on ncurses: motions, counts, search, undo, dot-repeat and syntax highlighting. ATLASRust A software 3D renderer from scratch. The only dependency is a framebuffer.

Smaller things: WikiRacer, Twisty, eclisp, Deep Double Descent, Trace, quack, SAT solver, tape, parser combinators, micrograd.ts, wtfdist, tabby, flow fields.

Hackathons

Hack the NorthBest Use of Gemini

Berkeley AI HackathonFinalist

Hack the 6ix3rd overall

Elsewhere

Off the keyboard I race gravel and ride mountain bikes. Grew up ski racing in Collingwood, Ontario.