General MotorsCase study · AI & Software
AI Electrical Design-Audit Platform
An AI platform that audits wiring, power, and network designs in minutes instead of hours or days.
- My role
- Sole developer: architecture, backend, frontend, AI features, testing, and deployment
- Timeframe
- General Motors · Oct 2023 – Present
- Team / context
- Internal platform available to all GM Automation engineers and suppliers
- Tech stack
- Python · FastAPI · React · PostgreSQL · pgvector · LLM APIs · RAG · YOLO · OCR · Single sign-on · Docker · CI/CD · Kubernetes · Azure
01
Overview
Before automation equipment is built, its electrical design has to be reviewed. Wiring, power, and network drawings are checked to make sure the system is safe, correctly sized, and complete. Done by hand, those reviews take hours or days.
I built an AI platform that does the review automatically. It reads electrical drawings with YOLO and OCR, audits the wiring, power, and network design, and flags problems such as power supply overloads and missing safety devices. It also includes an LLM and RAG assistant. I designed, built, tested, and deployed it as the sole developer, and it's available to every GM Automation engineer and supplier.
02
The problem & objective
Manual design reviews are slow. Auditing a set of wiring, power, and network drawings took hours or days of engineering time.
They're also where costly mistakes can be caught early, or missed. A power supply that's overloaded or a safety device that's missing is far cheaper to fix on a drawing than on built equipment.
The objective was to make those audits fast and consistent for everyone who designs automation for GM, including suppliers.
03
My role & responsibilities
Sole developer: architecture, backend, frontend, AI features, testing, and deployment. Delivered as part of my role as Controls Engineer at General Motors.
- Defined the platform's architecture end to end as its sole developer
- Built the Python/FastAPI backend and the React frontend
- Designed the PostgreSQL/pgvector data layer behind the LLM and RAG assistant
- Implemented drawing reading with YOLO and OCR
- Integrated single sign-on for engineers and suppliers
- Wrote and maintained a suite of 1,200+ tests
- Shipped the platform through Docker, CI/CD, and Kubernetes on Azure
04
Approach & process
Define the audits that matter
Focus the platform on the design reviews that consume engineering time (wiring, power, and network) and on the issues that matter most, such as power supply overloads and missing safety devices.
Teach the system to read drawings
Electrical drawings are visual documents. YOLO handles object detection and OCR handles text, turning drawings into information the platform can check.
Ground the AI assistant
Pair an LLM with retrieval-augmented generation (RAG), backed by embeddings stored in PostgreSQL with pgvector, so the assistant answers from retrieved context.
Build the application
A Python/FastAPI backend and a React frontend, with single sign-on so engineers and suppliers can use it securely.
Test everything
Build a test suite of more than 1,200 tests so the platform can keep changing without breaking what already works.
Ship and operate
Package the platform with Docker and deliver it through CI/CD to Kubernetes on Azure.
05
Key features & deliverables
Automated design audits
Wiring, power, and network designs are audited in minutes, flagging issues such as power supply overloads and missing safety devices.
Drawing reading
YOLO and OCR extract what's on an electrical drawing so it can be checked.
LLM + RAG assistant
An AI assistant grounded with retrieval-augmented generation over a PostgreSQL/pgvector store.
Enterprise access
Single sign-on, with access for all GM Automation engineers and suppliers.
Production delivery pipeline
1,200+ tests, Docker, CI/CD, and Kubernetes on Azure.
06
Challenges & how I solved them
| Challenge | How I solved it |
|---|---|
| Drawings are images, not data | I combined YOLO for detection with OCR for text, so the platform can work from the drawing itself. |
| AI answers need to be trustworthy | I used retrieval-augmented generation with pgvector embeddings, so the assistant works from retrieved context rather than the model alone. |
| One developer, an entire organization of users | More than 1,200 tests and an automated CI/CD pipeline let me change and ship the platform safely on my own. |
| Engineers and suppliers both need access | Single sign-on provides secure access for GM Automation engineers and suppliers. |
07
Results & impact
- Design audits that took hours or days now take minutes
- Catches issues such as power supply overloads and missing safety devices
- Available to all GM Automation engineers and suppliers
- Backed by 1,200+ tests and deployed through Docker, CI/CD, and Kubernetes on Azure
08
Skills demonstrated & lessons learned
Skills demonstrated: Full-stack development · System architecture · Applied AI (LLM, RAG, embeddings) · Computer vision (YOLO, OCR) · Testing and CI/CD · Cloud deployment (Kubernetes on Azure) · Electrical design domain knowledge
Domain knowledge decides what's worth building. I knew which checks mattered because I've done power and thermal calculations and drawn schematics by hand.
For a solo developer, a large test suite is what makes moving fast possible rather than risky.
Get in touch
Let's work together
I'm open to Controls Engineer, Automation Engineer, Software Engineer, and Full-Stack Developer roles, remote or on-site. If you're building something where automation and software meet, I'd like to hear about it.
elshark.atef@gmail.com