All work

Solo build

FORGE, AI engineering analysis platform

AI / software engineering · April to May 2026 · 6 of 8 phases built

Project

Discipline
AI / ML systems
Timeline
April to May 2026
Role
Built soloProduct design, architecture, full-stack development
Category
AI / software engineeringMultimodal AI platform for mechanical engineers
Status
6 of 8 phases builtLast commit 6 May 2026

Tools

  • React
  • FastAPI
  • Python
  • LiteLLM
  • RAG
  • Voyage AI embeddings
  • Supabase pgvector
  • PyMuPDF
  • Claude API
  • GPT-4o

An engineering question-answering app

FORGE is an engineering question-answering app: upload PDFs and drawings, then ask questions that are answered from them.

The workspace is a sidebar and a full-width chat: the engineer's file (a PDF or an image) and the AI's guided questions sit in the sidebar, and answers come back in the chat.

The core is a RAG (Retrieval Augmented Generation) knowledge base built from engineering PDFs. Every query is answered with retrieved chunks from it placed in the prompt.

Technical architecture

  • RAG knowledge base (Supabase pgvector): Documents are split into 500-word chunks with a 50-word overlap and embedded with Voyage (1,536 dimensions). Each question pulls the five closest chunks from a Supabase pgvector store.
  • Multi-model routing (LiteLLM): Calculation questions go to GPT-4o; images and everything else go to Claude Sonnet 4.5. The user can override the choice.
  • Multimodal input processing: Images are base64-encoded for the vision model. PDFs are processed with PyMuPDF. Text context is combined with file data.
  • Structured output layer: Reports include six sections: problem statement, approach, shown calculations, findings, recommendations, and a confidence level. Reports export as PDF. Session history is saved in the browser.
  • Guided onboarding per upload: The moment a file is dropped, the AI analyzes it and generates three targeted questions, such as load conditions, material or safety factor.

Not yet evaluated: there is no retrieval or answer-accuracy benchmark.

6 of 8 phases built · last commit 6 May 2026