Pool Takeoff Harness 🏊

AI harness for JJZ steel pools: reads pool construction PDF plans, extracts steel-takeoff dimensions with a local vision model, and fills the steel-bar calculator.

Quick start

start.bat            # installs deps if needed, runs the server, opens the browser

Then drop a PDF plan in the upload zone and click Extract Values with AI. Review the filled fields (especially Length/Width and spacings), then Calculate Steel Bars.

After extraction the UI shows which pages the AI read (e.g. "read page 3 of 21"). If it read the wrong sheet, type the correct page number(s) under 📄 Pages Read and click Re-extract from page(s) — handy when the dimensioned pool plan sits deep in a big set.

Requires the local vision model API running at the API_BASE_URL in .env (default http://192.168.1.127:8080, llama.cpp, OpenAI-compatible).

Validate against all plans

run-batch.bat                       # all PDFs in the plans folder
node scripts/batch.js --only Von Arx
node scripts/batch.js --limit 3
node scripts/batch.js --dir "C:\some\folder"

Writes results/<plan>.json (full values + calc + which pages were read) and results/_summary.json.

Accuracy feedback loop

In the web UI, after reviewing/fixing the extracted values, click 💾 Save as Correct. This stores the AI output vs your corrected values in results/feedback/. Then:

node scripts/accuracy.js     # per-field exact / within-5% / mean-error, AI vs your corrections

Use it to measure whether prompt changes actually improve accuracy over time.

Tests

npm test                     # locks the steel calculator math (lib/calc.js)

How it works

PDF → pick the most relevant pages (pdfjs text scan) → render with pdf-to-img → vision model → JSON (dimensions, spacings, auto-detected extra-bar features) → steel calculator.

Full details, the exact prompt rationale, and accuracy findings: METHODOLOGY.md. Session state / where to pick up: HANDOFF.md.

Gotchas (already handled, don't undo)

• The model is a reasoning model — requests must send chat_template_kwargs:{enable_thinking:false} or it returns empty content. (lib/extract.js)
• sharp cannot render PDFs on this machine — pdf-to-img does the rasterizing.

Config (.env)

API_BASE_URL, VL_MODEL, PORT. Render tuning: MAX_PAGES, MAX_SCAN, RENDER_SCALE, MAX_WIDTH, JPEG_QUALITY.