Team

From hand-made packs to a pipeline

The first asset packs were made entirely by hand: prompting in Stable Diffusion, editing and map generation in Photoshop, assembly and verification in Unity Hub, then submitting each pack to the stores manually. Those packs exist and went into distribution.

AIOS came out of the pain of that process. It is not a generator dreamed up at a whiteboard: it automates a procedure that had been clicked through by hand, by the same person who then wrote the code for it. Every stage of the pipeline maps to a step that previously had to be done by hand.

The work is done in a single orchestration session against a real task queue, with every code change going through a patch, a test gate and an explicit approval before it reaches the production tree. The same discipline applies to this website: what is claimed here is what the repository can be shown to contain.

What is done and what is being built

Done

• Hand-made asset packs — produced with the manual workflow above and published to the stores.
• Asset Factory pipeline — texture synthesis, tiling verification, PBR map computation and packing into a .unitypackage, in pure Python.
• Execution layer — SQLite task queue with time-boxed leasing, a patch harness with an executable test gate, and three-tier model routing under a budget guard.
• Operator panel — FastAPI backend with chat, dashboard, mail, social, projects, queue, database and staging tabs.
• This website — static files on Cloudflare Pages, no external CDN, three languages served by an in-house translation engine.

Being built

• The autonomous end-to-end pipeline — today the stages run, but a human still starts and closes the run.
• The scene dependency graph in Portable Animator — a specification and data model today, not a rendering engine.
• 4K upscaling inside the application — currently an external tool in the production workflow.

The automation does not push packs to the stores. Publishing remains a manual step and is described as such in the publish queue in the repository.

The AIOS technology stack

• Orchestration and queue: Python and SQLite in WAL mode. Tasks are claimed atomically (UPDATE … RETURNING in one transaction) and stamped with a time-boxed lease, kept alive by a heartbeat, recovered by an orphan reaper and restarted by a process watchdog.
• Execution layer (Harness v2): code changes arrive only as SEARCH/REPLACE patches, applied first to a staging copy. Promotion into the repository happens after a green test gate, with bounded self-correction — at most two attempts.
• Model routing, three tiers: Tier-0 runs local models through Ollama at zero API cost; Tier-1 is a cascade of free-tier providers (Cloudflare Workers AI, GitHub Models, Google AI Studio, Groq, OpenRouter); Tier-2 is paid cloud (Azure AI Foundry, AWS Bedrock) behind a budget gate.
• FinOps: the cost of every call is written to a ledger, with per-provider daily and total limits and a hard cutoff on breach. The guard fails closed if its state file is missing.
• Asset Factory: pure Python — texture synthesis, a wrap/offset tiling test, PBR maps (Sobel-filter normal, roughness, metallic, cavity-based ambient occlusion) and packing into a .unitypackage with hand-written .meta/.mat YAML, without launching the Unity Editor.
• Portable Animator: a local-first pipeline of script → images → voice → video assembly: story text and images default to the cloud (AWS Bedrock, Azure) with automatic fallback to the local pipeline, and everything else runs on the user's machine. The scene DAG is a specification and data model today, not a rendering engine.
• Observability: every model call leaves a JSONL trace — time, model, tokens, cost — read back by the panel.

The same layers are described in more detail, tier by tier, in the Engine section on the home page.