HIRE4GIG
A guarantee engine that re-estimates what each worker can realistically earn every three minutes, then matches orders to workers with a cost function that prices fairness in directly.
The problem
Platforms offer earnings guarantees to keep workers online during quiet periods. The naive version is a flat promise: work this window, earn at least this much. It fails in both directions. Set it low and it doesn’t retain anyone. Set it high and the platform pays out enormous sums to workers who would have hit the number anyway.
The guarantee needs to be personal, and it needs to move.
What I built
A Gaussian process regressor estimates a demand curve per worker, smoothed with an exponential moving average so a single quiet window doesn’t drag the estimate around. Every 180 seconds each worker’s guarantee is recomputed against their own recent demand rather than a platform-wide average.
Order assignment runs as a bipartite match using the Hungarian algorithm, with the cost function carrying a fairness term. Rather than always matching every worker to every order, it switches between k-nearest and full matching based on how much demand there actually is, which is where most of the runtime savings come from.
Results
Over a simulated 50,000 orders across 14,880 time windows and 80 agents:
- Handout ratio fell to 0.2% of platform cost, an 82% reduction against the flat-guarantee baseline.
- Gini coefficient across worker earnings settled at 0.11, 27% more even than baseline.
- Each matching window resolved in 38ms, leaving real-time operation comfortably in reach.
What I’d change
The evaluation is simulated. The demand model is fit to synthetic order flow, and a real platform’s demand has structure (weather, events, restaurant closures) that the simulation doesn’t contain. The fairness result is the one I’d most want to re-test against live data before claiming it holds.
