Logistics Optimization · shipped · 2026 · Client: Assessment — Innovacio Technologies
Computes the cheapest physically loadable pouch-to-container packing plan for tea export shipments.
Tea exporters decide pouch sizes, carton dimensions, pallet layouts, and container loading manually, producing low container utilization, avoidable freight spend, and packaging waste. Optimizing each stage in isolation fails structurally: a measured greedy pipeline achieved only 36.9% container utilization because each locally optimal choice constrained the next.
An exhaustive joint search enumerates ~15,000 complete pouch × carton × pallet × container configurations and scores each on total landed cost (material + corrugated board + freight), solving all stages as one problem in under a second. Savings are measured against an independently modelled 'conventional practice' baseline costed with identical physics, and plans respect real-world loading rules — 50 mm forklift clearance and crush-safe board grades. FastAPI/PostgreSQL backend with a Next.js frontend; the LLM never invents numbers — it explains results and answers what-if questions by calling the real optimizer via function calling.
The natural design is a stage-by-stage pipeline: pick the best pouch, then the best carton for that pouch, then the best pallet, then load the container. We built it and measured it: 36.9% container utilization. The failure was structural, not a tuning problem — every locally optimal choice quietly constrained the next stage.
Instead of optimising each stage, we enumerate ~15,000 complete pouch × carton × pallet × container configurations and score each one on total landed cost end to end. It sounds expensive; it resolves in under a second, and utilization roughly doubled to ~69%.
A savings number nobody trusts is worthless. We modelled a "conventional practice" baseline independently and costed it with identical physics, so every claimed percentage can be traced to a specific driver rather than asserted.
The assistant never invents figures. It explains results and answers what-if questions by calling the real optimizer through function calling — so the narrative always matches the maths.
Every stage is solved jointly rather than sequentially, then scored on total landed cost with real loading physics applied as hard constraints.
OpenAI GPT-4o-mini, Function calling, Python, FastAPI, SQLAlchemy 2.0 (async), PostgreSQL, Alembic, Next.js 14, TypeScript, Tailwind CSS, shadcn/ui, Three.js (react-three-fiber), Recharts, Docker Compose, GitHub Actions CI, Pytest, Playwright