Upgrading a live platform across 185+ sites from RFID to AI computer vision

Context

The platform is an AI vision and IoT system for access automation and asset tracking. It runs across 185+ live sites for clients including JSW, CWC (180+ sites), Landmark CFS, DHL and Haier, and it was originally built on RFID.

As Delivery Manager at Lynkit Solutions, I own end-to-end delivery of 2 enterprise SaaS products, along with an INR 5 Cr+ revenue portfolio.

Why we moved away from RFID

RFID worked, but it caused problems at the gate every day:

• Tags got lost, damaged or swapped between vehicles.
• Every site had to buy and issue tags, which added cost.
• Vehicles often arrived without a tag, so entry slowed down or had to be done by hand.
• Readers picked up the wrong tag, couldn’t read some tag types, and sometimes missed tags that were right there.
• Each tag had to be mapped to a vehicle number by hand, and mistakes crept in.

The challenge

Move the platform’s recognition from RFID to AI computer vision while it stayed live at every site, and add new capabilities on top:

• AI-based number-plate recognition
• AI-based ID recognition
• Damage detection

Every site had to keep running through the change, and every client needed to know what was changing and when.

My role

• Defined the requirements and scope of the upgrade.
• Coordinated engineering, QA and site teams.
• Managed client communication through the change.
• Owned escalations, change requests and SLA adherence across all live sites.

How we did it

• Pilot first. We started with one pilot site before rolling out anywhere else.
• RFID and AI side by side. At the pilot, both systems ran in parallel, so the gate never depended on the new system alone while we checked its accuracy.
• A small team of six: me, a product manager, an AI developer, a developer, a hardware engineer and a tester.
• Two months to implement it across all sites, from the pilot to the last rollout.

The hardest part

Number plates. Plate formats vary, and some characters look almost the same to a camera: 1 and I, Q, O and 0, M and W, V and Y. We annotated training images specifically for these look-alike characters, so the model learned to tell them apart.

Outcome

• No tags needed. Vehicles are recognised by their number plates, so sites no longer buy, issue or track RFID tags.
• 99% ANPR accuracy on number-plate reading.
• No more manual entry or mapping errors at the gate.
• Faster and more accurate entry for every vehicle.
• The client asked us to roll it out to all their sites and to move to new ANPR-based hardware.