Problem
A ride-hailing and delivery marketplace makes dispatch decisions continuously: send the nearest driver now or hold for a possible batch, how far to search, when to combine orders. These settings interact, and the right value for Friday at seven is wrong for Tuesday at two. Grab could test about a dozen strategies a year, because each needs weeks of engineering and weeks of production experiment. Some questions have no production answer, such as rerunning last Saturday with 20% fewer drivers.
Example from the article: a food order arrives from a mall while one driver is nearby and another is finishing a drop-off. Dispatch now, or hold in case a second order arrives within two minutes?
Solution in brief
Grab built sim-rs, an offline simulator that replays a city-day of historical bookings through the real dispatch logic in tens of minutes. It has swappable stages and a built-in A/B/n experiment engine. Coding agents run an autoresearch loop on it to find better strategies and to keep the simulator in line with production, under safeguards against metric gaming.
Questions this case answers
- How to evaluate a policy change in a two-sided system with interacting decisions when production experiments are slow and some counterfactuals cannot be run at all
- How to let an automated agent search for better policies against a simulator without it exploiting loopholes in the metric
- How to test a marketplace dispatch strategy offline before spending weeks on a production experiment?
- How to compare several strategy variants inside a simulator under the same demand and supply?
Tools and methods
- sim-rs — tick-based marketplace simulator: config and data in, one report out
- Switchbacks and spatial splits — A/B/n arms inside the simulator
- Survival-analysis cancellation models — decide which bookings leave the pool
- Autoresearch loop — agent proposes, tests, keeps or reverts a change
Google or a one-time email code. No password.