In a digital lending platform, applicants submit their PAN card as identity proof. The challenge: ensuring the name they enter actually matches the name on the PAN. Sounds simple. Indian names make it surprisingly hard.
The existing solution made two third-party API calls per submission — one to fetch the PAN holder's name, one to score similarity. It worked. But under marketing campaigns, every call had a cost that multiplied fast.
I noticed real-world name variations are limited: typos, transliterations, initials, reversed order. I hypothesised a weighted ensemble of string-similarity algorithms tuned for Indian names could handle most cases locally — at zero marginal cost.
Built a prototype, validated it against 30,000 historical records, presented to the architect and Risk Management. Got approved. Deployed with a fallback band — scores near the threshold still hit the external API as a safety net.
Result: significant reduction in external API calls on clear-match and clear-reject cases, cutting both latency and cost — while keeping the original API for ambiguous scores.
