WeldCount – AI-Powered Welding Symbol Detection Platform

Full-stack SaaS application that automates welding symbol detection from engineering drawings using computer vision and deep learning.

The Problem

Engineering firms spend hours manually counting welding symbols on blueprints for cost estimation. This manual process is error-prone, time-consuming, and doesn't scale.

The Solution

WeldCount processes uploaded PDF drawings through a GPU-accelerated ML pipeline (Detectron2 Faster R-CNN + Tesseract OCR) to automatically detect, classify, and count welding symbols with high accuracy. Users can review results in an interactive canvas-based viewer, make corrections, and export data for project planning.

Key Technical Achievements

Architecture & Scale

• Serverless GPU Processing: Integrated Modal.com for on-demand NVIDIA T4 inference, eliminating infrastructure costs while maintaining sub-5-minute processing times
• Separate Layer System: Immutable AI-generated data (OCR, detections) stored independently from user edits, enabling version control and reprocessing capabilities
• Direct S3 Uploads: Presigned URL pattern bypasses Next.js server for multi-GB file uploads, with storage quota enforcement (10GB per user)

ML Pipeline

• Detectron2 Faster R-CNN: Custom-trained object detection model (150 images) for welding symbol recognition
• Tesseract OCR: Text extraction with word-level bounding boxes for drawing annotations
• Batch Processing: Payment-gated processing with real-time webhook updates (PENDING → PAID → PROCESSING → COMPLETED)

Full-Stack Implementation

• Frontend: Next.js 15 with Turbopack, React-Konva interactive canvas, PDF.js rendering
• Backend: PostgreSQL (Prisma ORM), Redis caching/rate limiting, AWS S3 for artifact storage
• Auth & Payments: Clerk authentication with webhook sync, Lemon Squeezy (Merchant of Record) for global tax compliance
• Security: Redis-based rate limiting (20 req/min uploads, 5 req/min payments), input validation with Zod, storage quota enforcement

DevOps & Testing

• Comprehensive Test Suite: unit/integration/security tests with parallel CI/CD execution (fast vs. slow test separation)
• Local Webhook Development: Cloudflare Tunnel setup for testing payment/processing webhooks locally
• Type Safety: Strict TypeScript with zero implicit any, enforced via CI

Tech Stack

Next.js 15, TypeScript, PostgreSQL, Redis, AWS S3, Modal.com, Detectron2, Tesseract OCR, Clerk, Lemon Squeezy, Prisma, Tailwind CSS v4, Bun

Outcomes

• 75-80% Complete: Core ML pipeline, payment flow, and document viewer fully operational
• Partnership needed: Looking for an organization who would be willing to share collections of engineering documents to increase model accuracy and reliability.
• Production-Ready Infrastructure: Rate limiting, storage quotas, error tracking, and security logging in place
• Real-World Impact: Reduces manual symbol counting from hours to minutes with 90%+ accuracy