AssetForge

What it is. AssetForge is an AI asset-processing platform I built for production art and content pipelines. It wraps a fleet of image and 3D models — background removal, vectorization, image generation and image-to-image, upscaling, depth maps, 3D generation, STL-to-SVG paint masks, and music generation — behind one authenticated surface, and it exposes that surface four ways: a web dashboard, a REST API, a CLI, and MCP tools so agents can call it directly.

What it does. The dashboard is the operator's view: usage stats (jobs run, success rate, average time, active workers), a Workflow Execution Time chart plotting completed vs. failed across the last 50 runs, a Jobs-by-Model breakdown, and a live Job History side panel — plus a Playground for one-off jobs, a Workflows tab for chaining steps, an LLM Guide, and API Docs. The point is that whether a request comes from a designer clicking Playground, a script hitting the API, or an agent calling an MCP tool, it's the same job lifecycle underneath.

Why it works this way. Every model usually arrives with its own endpoint, payload shape, cold-start behavior, and polling quirks. AssetForge absorbs all of that once so apps and agents can just ask for an asset — the orchestration, retries, and status tracking are solved centrally instead of re-implemented per integration.

The interesting engineering. A FastAPI hub owns auth, API-key and DoubTech SSI login, usage telemetry, and the full async job lifecycle — async submit, sync submit, status polling, streaming, and cancellation — against a fleet of serverless GPU workers. PostgreSQL holds state with Alembic migrations, and each worker is built as its own GPU container from a shared base image, so adding a new capability is a new container against a known contract rather than a new bespoke service.

Stack: Python 3.12, FastAPI, SQLAlchemy + Alembic, PostgreSQL, React + Vite, Docker, RunPod serverless, MCP.

Links: assetforge.jaxns.net