Context
A media-planning platform had to allocate advertising spots across broadcast slots against rating targets — a hard combinatorial problem.
Task
Make solver strategies comparable and measurable instead of betting on a single heuristic.
Approach
Built three interchangeable engines (greedy, branch-and-bound, CP-SAT sidecar) behind one interface, with SoA data layout, SharedArrayBuffer worker pools, and SSE streaming.
Result
Honest benchmarking of exact vs. heuristic strategies on real workloads, with shared Zod/JSON-Schema contracts across Node and Python.
