Projects
Everything we build is one of two kinds: platform, which is general and open source, or an application, which is pointed at a domain and judged by whether it wins.
Platform
| What it is | Where it is | |
|---|---|---|
| Humanize 2 | The runtime. Orchestrate, execute and observe agent flows across ten coding-agent CLIs. | hmz.humanfia.ai · humanfia/humanize2 |
| flowverse | The official flows — the methods themselves, as directories of Python anyone can run or fork. | humanfia/flowverse |
| FlowBench | The benchmark that scores flows against each other on long-horizon work. In development. | Not yet released |
| oh-my-humanize | The workflow-native terminal coding agent we built first, and what it taught us. | humanfia/oh-my-humanize |
Applications
| What it does | Result so far | |
|---|---|---|
| KDA | Kernel Design Agents: research, implement, profile and iterate on performance-critical CUDA kernels. | Top-three track placements at the MLSys 2026 FlashInfer kernel contest |
| HOA | Humanize Olympic Agents: olympiad and competition mathematics, formally verified. | 6/6 at IMO 2026 · 670/672 on PutnamBench, ranked #1 |
How they relate
Platform is what we give away; applications are what prove it works. A flow that only ever wins on a benchmark we wrote ourselves has proved nothing, so the applications are chosen where somebody else keeps the scoreboard — a contest, a formal verifier, a public leaderboard.