Optical Character Recognition (OCR) Pipeline at Airtel ranks 6th Globally

Yes, we are talking about our optical character recognition (OCR) pipeline that extracts & processes customer information from scanned images of Know Your Customer (KYC) documents.

And yes, we participated in the ICDAR competition and ranked 6th globally.

So how did it happen?

Customer OnBoarding — Past & Present

Take an example of a person walking into a sim card store. You hand over a few documents, sign a few papers and voila! You get a new sim card activated by end of the day.

Ever wondered what happens in the background? Your First name, last name, father’s name, date of birth and address are extracted out of these documents, Know Your Customer (KYC) process is completed and your SIM card is activated.

Unique Identification Authority of India (UIDAI) had published an API that helped private companies complete user KYC in real time. Just Aadhar number and biometric match — and you get a sim card activated right away!

However, the Supreme Court verdict this past September barred private entities from using Aadhar information for KYC purposes. This practically served as a death knell to a lot of startups who leveraged Aadhar eKYC API for their day to day ops.

At Airtel, the third largest telecom operator in the world and 2nd largest in India, we on-board about 300,000 customers a day & we faced the heat too.

This verdict meant that we had to extract information from the documents manually. It takes 30 seconds to process each document manually. So 300,000 documents per day for 30 days count to 8.5 years of manual effort!

Our costs would skyrocket overnight.

Customer OnBoarding — Future

There was a pressing need to develop something that could quickly extract the information from the hundreds of thousands of KYC documents at hand.

Arjun Variar from our team has been relentlessly building robust reading algorithm for the past 18 months.

We quickly built a KYC parser on top of this work.

How does it work?

1. Take a picture of a KYC document.
2. Pass it on to our KYC-OCR API.
3. The API gives you back first name, father’s name, last name, date of birth and your address with 95% accuracy!

All of this takes less than a second.

Impact

Our KYC/OCR pipeline helps us:

1. Process 300,000 documents per day and 29,000 at peak hour.
2. Save 8.5 man years in manual efforts per month! (Each document takes 30 seconds to process manually: multiply that by ~300,000 for 30 days.)

Sitting from left to Right: Shardul Lavekar, Yash Sharma, Harisankar H. Standing from left to right: Arjun Variar, Parth Mudgal, Apurva JHM

The robust reading challenge — ICDAR

We wanted to see where we stand globally.

The Robust Reading competition hosted by the ICDAR (International Conference on Document Analysis and Recognition), is the largest competition of its kind for document analysis, and holds significant prestige in the field. The competition dates back nearly two decades and sees the best developers and innovators from over 100 countries tussle to rank on a highly-competitive leaderboard.

We submitted our algorithm to see where we stand and we are ranked 6th globally.

What next?

We want to achieve:

1. The top spot on the leaderboard.
2. The ability to process 2x volume with the same infrastructure.
3. Improve processing time per request.

We will be making our API public soon. Stay tuned for our next blog post where we explain how we built the robust reader.

Special thanks:

Arjun Variar, Parth Mudgal & Yash Sharma for making this happen.

Shrijata Mukherjee for helping us to put this together.