Archit Goyal
D-104MARK TYPE SIZE FIRE

Automation · 2025–2026

AI Take-offs

AI-assisted extraction of door and frame schedules from architectural drawing sets

A two-stage AI pipeline that reads architectural drawing sets, finds the door and frame schedules, and extracts every opening into a clean take-off. A lead pipeline runs it automatically and routes the best projects to estimators.

Context
Avalon International Aluminum
Role
Computational designer
Year
2025–2026
Stack
  • Next.js
  • TypeScript
  • Claude API
  • pdf-lib
  • PostgreSQL
  • Prisma
  • Notion API
  • Microsoft Graph

The problem

2–5 h

of manual take-off for every project an estimator looks at.

Target active effort per project
< 30 min
Extraction accuracy
68%
Automatic priority rank
A–D

Client work under NDA. Shown as system diagrams drawn for this site; no client data.

The problem

Before a door package can be priced, someone has to read the drawings: hundreds of sheets, looking for the few that hold door and frame schedules, then copying every opening into a spreadsheet. It takes hours, for every project.

The pipeline

Stage one asks the model a cheap question: which pages matter? Stage two builds a focused PDF from just those pages and runs four narrow extraction passes, one per element type. The answers are repaired, validated, flagged and exported to Excel.

Extraction

  1. Input

    • Drawing setPDF, split into chunks if large

  2. Stage 1

    • Page discoveryWhich sheets hold door + frame schedules?

  3. Stage 2

    • Aluminium frames
    • Aluminium doors
    • Wood doors
    • Borrowed lights

  4. Check

    • JSON repairRecovers truncated answers
    • Validation flagsTypes, tags, sizes, fire ratings

  5. Output

    • Excel take-off

The lead machine

The same pipeline runs without anyone uploading anything. New projects arrive from a construction leads database, their drawings are downloaded and taken off, and each is ranked A to D. Qualifying projects land in Notion for an estimator, with a SharePoint folder ready.

Automated lead pipeline

  1. Find

    • Leads databaseNew projects with drawing sets

  2. Fetch

    • Download drawings

  3. Read

    • AI take-off

  4. Rank

    • Priority A–DBy scope and fire rating

  5. Assign

    • NotionEstimator assignment
    • SharePointProject folder created

Target: from hours of reading to under 30 minutes of checking.