TopCoder Reimbursement Challenge — Top Performer
A black-box reverse engineering challenge: given 1,000 historical input/output examples from a 60-year-old travel reimbursement system with no documentation, no source code, and no formula—replicate it exactly. The system takes three inputs (trip_duration_days, miles_traveled, total_receipts_amount) and returns a single dollar amount. My solution ranked among the top performers in the competition and earned an interview invitation from 8090, an AI company based in Menlo Park.
The Challenge
Most machine learning problems give you labeled data and a known output space. This one gave you 1,000 (input, output) pairs from a system nobody alive fully understands—employee interviews hinted at quirks like a “5-day bonus,” receipt penalties near certain thresholds, and mileage tiers, but nothing was confirmed. The goal: produce a run.sh script that accepts the three inputs and outputs the reimbursement amount to within cents, in under 5 seconds, with no external dependencies.
Evaluation used four metrics: exact matches (±$0.01), close matches (±$1.00), average error, and a combined accuracy/precision score.
The Approach
Rather than trying to hand-craft rules from the employee interviews (which were contradictory and incomplete), I treated this as a supervised learning problem with heavy feature engineering.
16 engineered features derived from the 3 raw inputs:
| Feature Type | Examples |
|---|---|
| Raw inputs | trip_duration_days, miles_traveled, total_receipts_amount |
| Ratios | miles_per_day, receipts_per_day, receipts_per_mile |
| Polynomial | squared terms for each input |
| Log transforms | log(miles + 1), log(receipts + 1) |
| Cross terms | miles × receipts, duration × miles |
| Binary flags | is_5_day_trip, high_receipt_flag, low_receipt_flag |
The model is XGBoost (gradient-boosted decision trees), which excels at capturing the non-linear, threshold-based behavior that legacy business logic typically produces. The binary flags were directly motivated by the employee interview hints about special-case handling.
Results
The solution achieved high accuracy on the 1,000 historical cases and generalized well to the hidden evaluation set—well enough to be recognized as a top performer in the challenge. This performance led directly to an interview invitation from 8090, an AI startup in Menlo Park building enterprise intelligence tools.