Hire Balaji B. - Fine-Tuning Engineer

At Best Buy, I’m shaping the next generation of enterprise GenAI systems for supply-chain and logistics intelligence — taking large language models from experimentation to real production impact. I led the design and deployment of multi-agent LLM frameworks that operate across forecasting, fulfillment, and issue-resolution workflows. Using LangGraph + CrewAI, I built conversational agents that collaborate to plan inventory, trace shipment delays, and explain forecasts in natural language to operations teams. To make these systems scalable, I developed a knowledge-graph-driven RAG layer that fuses structured warehouse data with unstructured shipment logs — letting LLMs retrieve real-time context instead of static documents. These agents now serve as an internal “decision copilot,” powering over 10,000 daily logistics operations. On the modeling side, I combined Llama-based models (Llama-3, CodeLlama, Mistral) with transformer-based forecasting architectures (PatchTST, N-BEATS) to blend reasoning with prediction. This hybrid approach improved demand-forecast accuracy by 37 %, reduced planning time by 48 %, and drove $2.4 M in annual savings through smarter fulfillment automation. Beyond building, I’m exploring evaluation and self-reflection techniques for agentic systems — measuring not just accuracy, but how effectively the AI collaborates with humans in live environments.

At Optum, I worked at the intersection of machine learning and early-stage GenAI, building production systems that made healthcare data more intelligent, accessible, and actionable. This was a time when transformer-based models were just entering enterprise use, and I focused on adapting that research for real-world automation in claims and EHR processing. I experimented with early GenAI methods like encoder–decoder architectures (T5, BERT, ClinicalBERT) to summarize lengthy EHRs and prefill claim forms — improving turnaround time by over 60 % while ensuring HIPAA-compliant data handling. These projects gave me first-hand exposure to the transformer revolution and shaped how I now think about language reasoning in structured healthcare data. On the traditional ML side, I built risk and anomaly detection pipelines for eligibility verification and fraud reduction — combining gradient-boosted models, time-series analysis, and rule-based inference. We reduced false-positive alerts by 35 % and automated several manual review loops across claims workflows. I also designed predictive modeling systems for patient readmission, cost forecasting, and utilization patterns — implemented via PySpark, AWS SageMaker, and Airflow. These models formed the backbone of a larger operational ML framework with 99.8 % production uptime and fully automated retraining cycles. While my GenAI work at Optum was exploratory, it laid the foundation for the LLM deployment, RAG, and multi-agent systems I would later build at Best Buy — transitioning from research curiosity to enterprise-grade AI engineering.