Context
Sam's Club, a Walmart division, runs a supply chain spanning domestic trucking and international ocean freight. Fulfilment centres depend on real-time shipment tracking to plan labour and allocate docks, so the tracking platform is operationally load-bearing — when it reports a truck in the wrong state, people are staffed against a delivery that isn't coming.
Separately, distribution centres pack orders using automated boxing algorithms — greedy, brute force, and volume-ordered variants — chosen case by case. Which strategy wins depends on what is being packed, and that varies by centre.
Roles
Shipment Tracking Platform
Joined a team of ten owning Sam's Club's core tracking platform (Java/Spring Boot, MySQL, Kafka), and progressively took ownership of testing, event pipeline reliability, production monitoring, and a greenfield ocean freight initiative.
Backend development and production support: Built REST endpoints in Java/Spring Boot as the platform split from a single service into several dedicated ones. Set up XMatters alerting for geotagging failures, handled vulnerability remediation through Snyk and SonarQube, and ran live production investigations alongside fulfilment centre operations staff.
Integration testing rollout: Introduced the Karate framework across every team repository as a merge-blocking CI step, with four to five cases per endpoint. Presented the rollout to the wider organisation in a recorded sprint demo.
Kafka event deduplication: Truck state events arrived duplicated and sometimes contradictory. Built a Redis-backed deduplication layer suppressing repeats within a rolling window, with retry logic for failed writes.
Ocean freight tracking: Sole engineer extending the platform to international shipping — scoping the partner integration, defining the MySQL schema, and designing the target architecture. Analysing the partner's ETA model in Databricks showed a systematic bias near ports, which led to a more accurate model built on historical route data.
Boxing Algorithm Optimisation
Benchmarking and analysis: Compared Sam's Club's FASTr algorithm against alternatives used elsewhere in Walmart — a volume-based approach and an extreme-points heuristic — on execution time, volume utilisation, CPU, and heap. Analysis of real production orders in Python showed item size distributions varied substantially between distribution centres, and that no single algorithm was best everywhere.
Improved routing: Built an alternative decision tree routing orders to packing algorithms based on item characteristics, and simulated the full production order set against both trees — a roughly 5% gain in volume utilisation. Presented the findings to the staff engineer and engineering manager.
What I took from it
The bin-packing work started as a search for a better algorithm and ended somewhere else. No single strategy won everywhere, because the item size distributions differed enough between distribution centres that the right choice was a property of the centre rather than of the algorithm. The gain came from routing to the existing options well, not from replacing them.
The tracking platform taught the operational half of that. Duplicate and contradictory truck events were not an edge case to be tidied up later — they were what the system actually received, and building for them was most of the work.