SmartCon360's technology choices — and why

Stack choices give away a product's values. Here are ours:

Frontend: Next.js 15 App Router, React 19, TypeScript strict (no any), Tailwind v4, Prisma 6, zero-dependency SVG charts with D3. One application, on Vercel.

ai-service (Python): real engineering libraries on FastAPI — ifcopenshell (geometry + quantities from the IFC model), PyMuPDF (PDF), ezdxf (CAD). These are not "AI"; they are deterministic parse engines.

AI layer: Gemini 2.5 Flash (planning/reporting text), with Gemini 2.0 + Claude sonnet-4-5 as fallback on the vision side. The critical principle: quantities never come from an LLM — quantity take-off comes only from ifcopenshell geometry. AI is a suggestion tool, not the center of the product; every output also has a manual path.

Data & deploy: Neon Postgres (13 schemas), a single canonical Prisma schema; services on Railway.

An honest boundary: we don't use an off-the-shelf UI library like shadcn — we have our own design-token system, which means a maintenance burden.

Where do you think the line between "suggestion" and "decision" should be drawn for AI in construction?