Manufacturing AI Design Review Engine
About this project
Manufacturing Design Review Engine
The Manufacturing Design Review Engine is an offline-first, privacy-focused manufacturability analysis system built for mechanical and manufacturing engineers working in industrial environments.
This project originated from a recurring challenge in product development: engineers need to evaluate whether a part design can realistically be manufactured — and by which process — early in the design cycle. Today this evaluation is often done manually, relying on the experience of a DFM specialist reviewing CAD geometry against process-specific rules. The process is slow, inconsistent, and difficult to audit.
The system addresses this by encoding manufacturability knowledge into a deterministic scoring engine. Given a STEP CAD file, the engine extracts geometry features — bounding box ratios, thin-wall presence, turning symmetry, sheet metal heuristics, extrusion profiles — and evaluates them against rule sets for eight manufacturing processes: CNC machining, CNC turning, sheet metal fabrication, extrusion, additive manufacturing, casting, forging, and MIM.
A key capability is AUTO mode: geometry-driven process classification that recommends the most suitable manufacturing process based purely on CAD analysis, without user bias. The classification is deterministic and regression-tested through golden test suites.
AI is present in the system but intentionally scoped. A local LLM — Ollama running Jamba2-3B — is used only for narrative explanation and report phrasing, never for scoring or classification decisions. This distinction is architectural: engineering decisions remain deterministic and auditable, while the explanation layer makes findings readable.
The entire system runs offline. FAISS vector store, local sentence-transformer embeddings, and local LLM inference operate without any cloud dependency, making the tool suitable for privacy-sensitive industrial environments where data cannot leave the network.
The result is a system that takes a CAD file and produces an audit-ready DFM report — including process scores, risk flags, material modifiers, and engineering findings — in a fraction of the time a manual review would require.
Tech stack: Python 3.11–3.13 · STEP geometry parsing · FAISS · Sentence-Transformers · LangGraph · LangChain · Ollama · Streamlit