Unpaid bills. Months of mounting charges with no payment in sight. Every month, another $5-6 million piling onto the tab. Sales reps were scrambling to appease furious customers who refused to pay for rides they couldn't verify. The CFO was asking "what the hell is going on?" And my boss looked at me and said: "Make this go away."

Quick Summary

I built a systematic fraud detection and classification system that could analyze thousands of healthcare rides line-by-line, separating legitimate trips from driver abuse, and gave clients the confidence to pay their bills by showing them exactly what they owed and why.

Problem

Lyft's healthcare business was built on a service called Concierge—dispatchers at health organizations could order rides for patients who didn't have the Lyft app. But this created a massive vulnerability: drivers controlled ride start/stop with no passenger oversight. Some drivers would complete legitimate trips then drive in circles for hours. Others would pick up patients, drive them home, and leave the app running overnight—we had one $975 charge for a 9-hour "ride."

The clients had spreadsheets with thousands of rides but could only spot-check obvious fraud manually. They were refusing to pay entire invoices because they couldn't trust any of it. We were duct-taping together explanations with incomplete data while bills kept mounting.

Approach

I realized this wasn't just a fraud problem—it was a data classification problem at scale. We had incredibly granular GPS data (lat/long coordinates captured every few seconds) that clients didn't have access to. The key was mapping every ride's actual path against its intended route and building systematic rules to classify what happened.

Working with analyst Kevin Danser, we mapped out the entire ride ecosystem: legitimate trips, gray-area scenarios (like hospital campus navigation), and clear fraud patterns. The constraint was that we needed to process thousands of rides quickly while being defensible to skeptical clients.

What We Built

A SQL-based classification system that could analyze every GPS coordinate in a ride and automatically categorize it:

Legitimate Rides: GPS showed the car reached within 0.15 miles of the destination (accounting for large hospital campuses)

Start-Stops: Driver confusion, minimal charges we'd absorb

Excess Mileage: Car reached the destination but continued beyond—we'd bill for the legitimate portion and write off the excess

Complete Fraud: Car never went to the destination—full writeoff

The system would take a raw invoice (1,500 rides, $1.2M) and output a reclassified file with adjusted totals and detailed explanations. But the hardest part was the storytelling: I created visual presentations showing actual ride routes on maps, pulling specific examples to walk clients through each classification with GPS evidence.

Impact

We recovered and processed hundreds of millions in previously uncollectable bills. That initial $100M backlog got resolved, and clients started paying monthly invoices again. The system became our standard process—every healthcare invoice got pre-processed before sending to clients.

More importantly, it shifted the relationship from adversarial to collaborative. Instead of clients saying "your drivers are scamming us," we could show them: "Here's the legitimate 90%, here's the 10% we're writing off, and here's exactly why." We even uncovered cases where dispatchers were collaborating with specific drivers to game the system.

Closing Thought

This was the moment I learned that good fraud detection isn't about catching bad actors—it's about creating transparency that everyone can trust. Sometimes the most powerful tool isn't preventing the problem, but being able to explain exactly what happened.