Problem
A leading logistics marketplace relied on a rule-based heuristic to rank moving service providers. The formula had grown complex and opaque, often suppressing high-quality providers and depressing click-through rates. The open mandate was to apply ML/AI to lift CTR and deliver a revenue uplift target of 5%.
Approach
We analyzed 10 years of activity and selected 4 post-pandemic years to reflect current behavior. Provider operations data and customer feedback were unified from three sources into a single training dataset. Two complementary ranking families were pursued in parallel: a deep-learning model and a gradient-boosting ranking model. With the client in the loop, we iterated more than seven versions of each and benchmarked them against the heuristic baseline.
Solution
- Production-ready ranking service exposing a consistent API for search results.
- Feature management via a managed feature store and a cloud data warehouse to keep offline/online parity.
- Containerized model serving with workflow orchestration for training, evaluation, and rollout.
- Continuous evaluation pipeline with A/B testing hooks and automated regression checks.
- Centralized logging, metrics, and alerting for reliability and rapid troubleshooting.
Impact
- Higher CTR versus the heuristic baseline in controlled online tests.
- Revenue uplift trending toward the 5% target, driven by better provider–customer matches.
- Improved marketplace fairness by ensuring consistently high-performing providers ranked prominently, while also giving new service providers a fair opportunity to gain visibility and customer trust.
- The success of this engagement led to additional projects, including ML models to predict the number of workers and job hours required based on job details, further improving operational efficiency and planning.