From Camera to Cloud: Deploying Our AI Queue Management System

January 4, 2026 (3mo ago)•3 min read
From Camera to Cloud: Deploying Our AI Queue Management System
#tech
#ai
#ml
#azure
#deployment

From Camera to Cloud: Deploying Our AI Queue Management System

This project started as a bold idea from my teammate Pragya Singh, who built the core of the system end‑to‑end. I joined in later to help debug, fix deployment issues, and ship it to the cloud. This post is a quick story of what we built, how I deployed it, what went wrong along the way, and what I learned.

What the Project Does

The system tackles a simple pain point: queues are frustrating. It watches queues in real time, predicts what’s coming next, and suggests how staff should be allocated.

example

The stack includes:

  • YOLO queue detection (camera → people count)
  • LSTM prediction (15‑minute forecast)
  • OR‑Tools optimization (best staff allocation)
  • Dashboard + notifications (live view and alerts)

Architecture in One Breath

  • AI Engine: Python + Flask (YOLO, LSTM, OR‑Tools)
  • Backend: Node.js + Express + MongoDB
  • Frontend: Next.js + TypeScript
  • Database: MongoDB Atlas
  • Deployments:
    • Backend + AI on Azure Container Apps
    • Frontend on Vercel

My Role

Pragya built the system. I came in to fix edge cases, resolve deployment issues, and get everything running in Azure + Vercel. That meant:

  • resolving container port mismatches
  • fixing dependency conflicts (TensorFlow + OR‑Tools + NumPy)
  • slimming down the AI container image
  • correcting backend → AI routing
  • stabilizing MongoDB startup

Deployment (Where It Lives)

  • Backend: Azure Container App
  • AI Engine: Azure Container App
  • Frontend: Vercel

Azure Container Apps handled ingress + scaling, while Vercel made frontend deployment simple.

Challenges I Faced

1) Port mismatches

The containers were listening on one port while Azure expected another.
Fix: set explicit PORT and AI_ENGINE_PORT.

enter image description here

2) Massive AI image

The AI container hit multi‑GB (more than 7GB) size because of ML dependencies.
Fixes: tensorflow-cpu, opencv-python-headless, tighter pins, .dockerignore.

3) Dependency conflicts

TensorFlow and OR‑Tools required incompatible protobuf versions.
Fix: pinned compatible versions across the stack.

4) Backend couldn’t reach AI

Internal DNS wasn’t resolving, so the backend couldn’t call the AI engine.
Fix: point backend to the AI engine’s external FQDN.

enter image description here

5) Mongo connection timing

Requests hit before Mongo connected.
Fix: wait for DB connection before starting the server.

What I Learned

  • Most “cloud problems” are actually configuration problems.
  • AI deployments are heavy — container size needs planning.
  • Explicit env vars prevent hours of guessing.
  • Working with a strong teammate makes complex projects possible.

Final Thoughts

This system is largely Pragya’s work — she built the AI pipeline, backend logic, and frontend experience. I’m proud of the role I played in making it stable and deployable.

If you’re building something similar, my biggest advice is: design for deployment early, because the last mile is often the hardest.

LinkedIn