Rajash Bordoloi

Built two production LightGBM forecasting models for H3 cell and area level demand, owning the modeling, validation, and serving lifecycle.

Deployed

TiRex for near real-time cell demand forecasting, reducing inference latency by two orders of magnitude through ONNX Runtime and multithreading. Built evidence-calibrated ML promotion gates with immutable artifacts and fail-closed deployment.

Machine Learning Engineer at Dispatch Network (2026-03 – Present)

• Built two production LightGBM forecasting models for H3 cell and area level demand, owning the modeling, validation, and serving lifecycle.
• Deployed TiRex for near real-time cell demand forecasting, reducing inference latency by two orders of magnitude through ONNX Runtime and multithreading.
• Built evidence-calibrated ML promotion gates with immutable artifacts and fail-closed deployment.
• Built a demand-first rider repositioning system using H3 spatial search and Redis-backed serving.
• Architected a dispatch decision engine and off-policy evaluation framework, enabling reliable policy evaluation through propensity tracking and immutable reward/decision logs.
• Designed a driver-centric semi-MDP with temporal discounting and an FQE→IQL path for offline policy evaluation.
• Evaluated dispatch policies in a service-in-loop simulation twin, showing +15.9% deliveries, +18.1% OPH, and +27.8% EPH offline.
• Drove ML technical direction, evaluating MLE candidates, mentoring interns, and reviewing model and system architecture.

M.Tech in Geoinformatics – Indian Institute of Technology Bombay (2024-01 – 2026-01)