whoami
Building has always been the way I make sense of things.
These days most of what I build is AI agents and the tools around them. I work on the boundary between model behavior and system behavior: what the model gets wrong, what the harness makes worse, and what can actually be made reliable.
Before agents, I spent years on backend systems where a bug costs real money, from payment callbacks at a fintech in Tehran to in-game currency services at Xsolla.
This is where I leave notes, experiments, findings, and occasional opinions from the systems I'm pulling apart or putting together. Mostly LLM harnesses, local deployments, open source, ways of working, and the occasional strange connection between otherwise unrelated fields.
selected
Sends one brief to several AI models for images, voices, sound effects and 3D, and records each pick with who made it and what it beat.
A small task ledger and finish gate for parallel Claude Code and Codex sessions.
ste-evalwhat it foundwhat almost fooled it
The harness behind the two eval posts, with 224 replies from Claude Code and Codex scored blind by judges from both families.
MSc researchComputer Communications 2024ACM DIVANet 2023
Six object detectors benchmarked on a Raspberry Pi 4 and a Coral Edge TPU, in two first-author papers.
I live in Vancouver, BC. If any of this touches your own work, send me a note.