Research and open source
Released tools, experiments and publishing workflows. An open-source label here always means a public repository with an applicable license. Findings ship with their limitations.

Released tools
OpenNPU
Local AI on the NPU of AMD Ryzen AI laptops, on Linux: chat, vision, voice and generation with its own open NPU engines.
Apache-2.0 license. Updated 2026-10.View repositoryLLM Cost Watchdog
Cost monitoring for AI agents. Tracks pricing across providers, enforces budget ceilings and detects runaway loops before they run overnight.
MIT license. Updated 2026-04.View repositoryCLI Usage
A desktop tray indicator that shows Claude Code and Codex usage limits at a glance.
MIT license. Updated 2026-10.View repositoryExperiments
Local inference on consumer NPUs
OpenNPU is the published part. Related benchmarking of tokens per second per watt and a custom matrix-multiply engine for the same hardware exist as working code but are not yet public.
Research. Numbers will be published with exact test conditions or not at all.Agent-driven playtesting
Scripted agents played and reported on Cellblock campaigns to find bugs and pacing problems before human review. Reports are kept with the project, not published.
Research, internal. Limitations: one game, one reviewer rubric.Publishing workflow
Radar
A Persian-first site where agents write a daily AI digest and a second agent checks every item against its source. The site shows a stale notice when the last run is old.
Published. Machine-written, agent-reviewed, no human editor on each item.Visit the site