Reconstruct AI

• 

  Landing Page

Inspiration

Watching how a real government tender actually gets bid on is what set this off. A drawing gets released, and before a contractor can even put in a number, someone has to spend 15–20 days manually reading every sheet — tallying rebar schedules, cross-checking dimensions across plan and section views, chasing current material rates city by city. That delay isn't just inefficiency, it's often the difference between making the bid window or missing it entirely. We wanted to know: what if the drawing itself could become the interface, and an AI pipeline could do the reading, measuring, and pricing — while staying honest about what it actually knows versus what it's guessing?

What it does

Reconstruct is a drawing-first workspace for construction estimation. Upload a tender PDF and it renders every sheet, indexes it into structural elements (footings, columns, beams, rebar schedules, notes), and lets you ask any element a question grounded in that specific part of the drawing — never a detached chatbot answer. It then reconstructs the whole drawing as an interactive, measurable 3D model, with every dimension traced back to a figured value on a real sheet, so someone without formal engineering training can click a beam and see its material grade and source sheet. Finally, a five-agent pipeline — vision reading, sandboxed quantity arithmetic, live ₹ rate search, risk/quality supervision, and a deterministic Sheets writer — reads every element, prices it against current Indian material rates, and files a complete Bill of Quantities directly into a Google Sheet the team already works in.

How we built it

The frontend is a Vite + React + TypeScript workspace where the blueprint is the home screen — a single stateful page owns the loaded document, the selected sheet/element, and the sheet-vs-3D view toggle, with panels for element understanding, chat, and BOQ estimation hanging off it. Three.js powers the Truth-3D viewer, mapping each reconstructed solid back to a bounding box on the original PDF page so provenance is never lost. On the backend, a FastAPI service renders and persists a structured element index per sheet, then streams a five-agent estimation run over SSE so the frontend can show live agent status as quantities get computed, priced, and written out. Quantity arithmetic runs in a sandboxed, deterministic layer, deliberately separated from the LLM-driven reasoning around it, so the final BOQ is reproducible rather than a different number every run. A Google Sheets integration, guided through a connect-sheet flow with a shared service account, is the delivery surface — because that's where estimation teams already work, not a new tool they have to migrate into.

Challenges we ran into

• Grounding accuracy under ambiguity — getting the pipeline to refuse to guess a missing dimension, rather than confidently hallucinating a plausible one, was harder than extracting dimensions that were clearly stated.
• Multi-agent determinism — keeping the final BOQ reproducible and auditable despite LLM-based agents in the loop, by drawing a hard boundary between "LLM judgment" and sandboxed code that must not vary.
• 3D provenance on messy scans — mapping every reconstructed 3D part back to its exact source dimension had to survive real government tender scans: low-resolution, rotated, and inconsistently scaled, not clean CAD exports.

Accomplishments that we're proud of

Getting a real government tender drawing to go from upload to a costed, Google-Sheets-delivered BOQ in minutes — with the model able to say "this dimension isn't on the drawing" instead of inventing one — was the moment the project felt real rather than a demo trick. The 3D provenance jump, where clicking any part of the reconstructed model takes you straight back to the originating sheet, turned out to be the feature non-engineers responded to most.

What we learned

That the hard part of applying AI to construction estimation isn't reading a drawing — it's knowing when not to answer. A wrong BOQ line costs someone a bad bid, so the "zero invented numbers" constraint had to be a first-class design principle from the start, not a guardrail bolted on afterward. We also learned that keeping deterministic computation separate from LLM reasoning is what makes an agentic pipeline trustworthy enough for a domain where the output has real financial consequences.

What's next for Reconstruct

Expanding the indexed drawing formats beyond PWD/CPWD conventions, building a proprietary rate database that improves with every processed tender, and piloting with a handful of EPC contracting teams to stress-test accuracy against their actual bid outcomes.

Built With

• fastapi
• openai
• react
• render