Marketplace and on-demand service platforms that scale.
Three systems that must work together under load: a matching engine, live dispatch, and payments with ratings. The complexity is in the multi-sided nature — you are balancing supply and demand in real time, and the failure modes differ on each side of the market.
Peak behaviour is designed for rather than discovered. That means knowing which parts of the system queue, which degrade gracefully, and which must stay real-time. Launching without that work is how marketplaces fail publicly on their busiest day.
Prediction helps, but the harder problem is fairness and transparency rather than the model. A pricing system users perceive as arbitrary damages retention on both sides of the marketplace, so the rules need to be explainable before they are optimised.
Retention is a product problem more than an algorithmic one: consistent earnings, fair matching and a dispute process people trust. Incentive and loyalty mechanics help, but they do not compensate for matching that providers experience as unfair.