Manas Kumar Jha

The problem

Banks caught fraud only after onboarding, through transaction monitoring. By then the mule account was already active and the damage done. The gap was pre-onboarding, catching a mule account before it entered the system.

What I did

Mapped the problem from scratch, wrote the PRD and detection logic for two typologies, shell-company detection across GSTIN, PAN, ROC, and Udyam, and individual mule detection through behavioral enrichment and digital-footprint analysis.

The hard call

Every bank wanted different weightage per onboarding journey. A fixed model meant a custom rebuild per client, so I designed a configurable scoring engine, each bank tunes its own risk weightage without touching core logic.

Outcome

Live in production with enterprise banking clients, shipped with Engineering, Sales, and Compliance.

The problem

Enterprise clients needed structured, reliable data out of unstructured documents for identity and compliance checks, without slow manual review, but AI extraction that's highly accurate is still wrong sometimes, and in a compliance workflow a confident wrong answer is worse than a slow right one.

What I did

Own Signzy's LLM-based document extraction and face-liveness APIs. Wrote the PRDs and extraction requirements, defined what counted as accurate enough for each document type, and partnered with engineering on prompt design and model evaluation.

The hard call

Could have optimized purely for speed and auto-approved everything the model returned, but a wrong answer in a compliance workflow costs more than a slow one, so set a higher accuracy bar for auto-accept and routed uncertain cases to human review instead.

Outcome

Live in production, turning unstructured documents into structured, decision-ready data for enterprise clients.

The problem

Signzy's EPFO verification flow had a long-running tail of production bugs: pre-production and production environments were returning differently structured responses, and a regex mismatch kept turning that into live incidents for enterprise clients.

What I did

Took ownership of the EPFO revamp. Raised 20+ production bugs myself, and drove the resolution of 90+ bugs and 7+ major incidents working closely with engineering.

The hard call

Rather than patching each bug as it surfaced, pushed for a systematic comparison of pre-prod and production responses end to end, which surfaced the regex mismatch as the root pattern behind most of the recurring incidents.

Outcome

Now live with 8-9+ banks and NBFCs on the revamped flow. Added multiple fallback integration paths, cutting 'data not found' failures, and holding a 98% success rate, even counting cases where the EPFO government site itself was down. Latency dropped significantly across the flow.

The problem

Enterprise clients needed to verify both individuals and organizations, employee and employer checks, blue-collar and white-collar, but that meant stitching together multiple separate verification APIs for each check type.

What I did

Built Signzy's unified Background Verification platform from scratch, bringing employee and employer checks, including EPFO-based employment verification, together into a single configurable API, so enterprise clients could verify both individuals and organizations through one integration instead of many.

The hard call

Could have kept employee and employer checks as separate flows per client, but chose to unify them into one configurable API so enterprise clients run a single integration instead of stitching together many.

Outcome

Live in production, used by enterprise clients to verify both individuals and organizations through a single integration.