Solo build — founder & lead engineer
FORGE — AI engineering analysis platform
Building the AI layer for engineering
Mechanical engineers today lose hours to fragmented toolchains — switching between CAD software, FEA tools that cost thousands per license, reference manuals, generic AI chatbots that don't understand engineering, and spreadsheets for manual calculations. FORGE AI collapses that entire workflow into one intelligent workspace.
The platform is organized around a split-screen workspace: the engineer's file (sketch, CAD, or PDF drawing) lives permanently in the left panel, always visible. The AI chat interface, guided questions, and structured output live in the right panel. Annotations from the AI — highlighted stress concentrations, traced load paths, flagged weak points — appear directly overlaid on the file in the left panel. Engineers see exactly where issues are, not just a description of them.
The core intelligence is a proprietary RAG (Retrieval Augmented Generation) knowledge base built from FEA literature, material standards, and engineering handbooks. Every query is answered using retrieved engineering knowledge — not general AI training data. This is FORGE AI's primary competitive moat: a purpose-built engineering knowledge base that gets more accurate as more domain data is ingested.
Three product layers
- Layer 1 — Build from scratch: Describe an idea in plain language. The AI generates a full concept: design parameters, material suggestions, initial dimensions, and a bill of materials. Fast ideation with engineering intelligence baked in.
- Layer 2 — Upload and collaborate: Upload an existing sketch, CAD file, PDF drawing, or image. The AI analyzes it, asks targeted guided questions about intent and constraints, then collaborates to refine and improve. Like having a senior engineer review your work in real time.
- Layer 3 — Structural analysis: Once a design exists, run it through physics-based analysis — stress concentrations, load paths, weak points, failure prediction, safety factors. Outputs show all work so engineers and their managers can verify every calculation.
Technical architecture
- RAG knowledge base (Supabase pgvector): Engineering PDFs are chunked, embedded as vectors, and stored in Supabase pgvector. Queries are vectorized and matched to the most semantically relevant engineering knowledge before the AI responds.
- Multi-model routing (LiteLLM): Smart routing assigns tasks to the best model — Claude for multimodal sketch interpretation, GPT-4o for analytical reasoning, Gemini Flash for fast Q&A. Engineers can override in settings.
- Multimodal input processing: Images and sketches are base64-encoded for vision models. PDFs are processed with PyMuPDF. CAD files are parsed with ifcopenshell. Text context is combined with file data for complete understanding.
- Structured output layer: Cited reports include summary, methodology, shown calculations, findings, and recommendations. CAD-ready parameters are exportable. Full iteration history is saved in Supabase so engineers can compare every version of their design.
- Guided onboarding per upload: The moment a file is dropped, the AI analyzes it and generates targeted questions — load conditions, material, safety factor, operating environment. This ensures rich engineering context before any output is produced.
Building FORGE AI from scratch required me to think like both an engineer and a product architect. The domain knowledge I've accumulated — from restoring an E30 to designing a medical device at WPI — is embedded in every product decision. I'm not building a general chatbot and pointing it at engineering problems. I'm building a tool that thinks the way engineers think, because I am one.
In development — 2026