May 2023 – Present · 3 yrs · Gurugram

• Multi-Agent LLM Planning System
• Deployed a multi-agent supply planning platform (LangGraph-based) used by 2,000+ daily users across trading & supply teams.
• Designed retrieval-grounded reasoning pipelines with rule constraints, reducing critical planning errors by 32% (human audit validated).
• Fine-tuned a 7B open-weight LLM using QLoRA (multi-GPU DDP) on domain planning data, reducing hallucination rate by 40% and deployed via AWS EKS (vLLM) thus reducing inference cost by 28%.
• Defined production guardrails, reducing hallucination rates, improving grounding and business trust metrics, and lowering P95 latency from 16s to 6s, increasing daily query capacity.
• Demand Forecasting & Supply Optimization (PyTorch Forecasting)
• Developed Temporal Fusion Transformer (TFT) models across 50+ countries, improving 26-week RMSE by 18% vs XGBoost baseline.
• Designed probabilistic multi-horizon forecasts optimized for service-level constraints (not just RMSE), reducing stock-outs and excess inventory.
• Productionized distributed training on SageMaker, reducing retraining time from 6 hrs → 4 hrs.
• Tracked parallel experiments using MLflow, enabling reproducible model training and hyperparameter tuning across 50+ countries.
• Led cross-functional roll-out across supply, trading, and merchandising teams.