Logistics Simulation & ML Pipelines

The Problem

When a logistics network misallocates resources, real shipments are delayed and real businesses are affected. Global logistics operations involve complex resource allocation with massive combinatorial search spaces, and traditional approaches couldn't adapt to real-world constraints. Kunumi was applying RL and NLP to serve logistics customers better, but experimentation cycles were slow and the path from research prototype to production was unclear.

What I Did

Built both the simulation layer and the ML pipeline infrastructure — the full path from research environment to production-ready models.

• Created simulation environments using OpenAI Gym that modeled real logistics scenarios with enough fidelity for RL agents to learn transferable policies
• Engineered ML pipelines on GCP for both Reinforcement Learning and NLP systems, focusing on reproducibility and iteration speed
• Architected experimentation pipelines for managing high-dimensional operational data, designed to reduce the feedback loop between hypothesis and result

The Hard Parts

Simulation fidelity is the core tension in RL for logistics. Too simple and the learned policies don't transfer — real customers experience the same inefficiencies. Too complex and training becomes intractable. Finding the right level of abstraction — one that respects real operational constraints while remaining tractable — was the most important engineering decision. Every improvement in simulation fidelity translated directly to better resource allocation for actual logistics operations.

Impact

• Improved resource allocation efficiency by 20% for global logistics operations
• Reduced experimentation iteration time by 30% through pipeline architecture
• Established ML pipeline patterns adopted across the engineering team