Team project. Built at the Black Forest Hackathon 2026 by the team hackhog. The repo belongs to the team. This page covers what I worked on: 38 merged pull requests.
What it is
In manufacturing, parts often arrive as a PDF drawing. Before that becomes an editable CAD model, someone redraws the part by hand. drawing2cad hands this to an agent: it reads the drawing, builds the model exclusively through the tools of an MCP server, and delivers a validated STEP model plus a dimensioned DXF/PDF drawing. Anything it can’t derive with confidence goes to a human as a review flag.
My contribution
- Agent loop in the web backend: start runs from the web UI, keep the step budget across model rate limits, shorten long tool results in later requests, timeouts per model turn instead of per request.
- Error messages for the agent: PDF errors go back to the agent as readable text, and an aborted run names the actual blocker. The agent receives the drawing references and has to deliver a complete drawing.
- Model profiles: manage models and endpoints on a settings page, including the Responses API. Each run records model, MCP version, reasoning effort and skill version, so runs can be compared.
- Run timeline: a run’s model turns and tool calls as a step timeline.
- Review flags as feedback: users resolve open flags, and their answers are collected as feedback. Plus measuring review time and manual effort.
- E2E tests: the web UI runs in Chrome against the real backend.
- Web UI: rebuilt in React with TanStack Router/Query and shadcn, plus a 3D view of the released STEP model.