Model Training, Fine-Tuning & Alignment

• Built an end-to-end fine-tuning pipeline for open-weight models (Gemma 2, Qwen2.5, Llama) using LoRA/QLoRA on curated golden datasets
• Engineered an "online golden-dataset" flywheel — production traffic → review → curated SFT/preference pairs → scheduled retraining
• Applied preference-based alignment (DPO/RLHF/RLAIF) to steer models to domain tone and compliance constraints
• Stood up an evaluation harness — golden-set scoring, LLM-as-judge, red-teaming — as a production promotion gate

Model Compression & Efficient Serving

• Led model compression — quantization (INT8/INT4, GPTQ/AWQ/GGUF), distillation, pruning — to fit larger models onto smaller GPUs (L4/A10G vs. A100-class)
• Optimized throughput with vLLM/TensorRT-LLM, paged KV-cache, continuous batching, and speculative decoding
• Ran a right-sized GPU serving cluster on AWS ECS with autoscaling across training and inference workloads

Self-Hosted Inference — Cost & Compliance

• Replaced third-party API inference with fully self-hosted, in-VPC models to meet SEBI data-residency requirements
• Centralized model access via a LiteLLM gateway — unified auth, routing, rate limiting, cost attribution, audit logging

Applied GenAI Products

• Architected a proprietary GenAI Agent Platform on Amazon Bedrock (Claude 3.5) with MCP — 98.7% less context overhead, 60% of support workloads automated
• Shipped a RAG assistant (LangChain/LangGraph); built an AIOps self-healing platform cutting incident RCA time ~92% (2 hrs → <10 min)
• Automated compliance ops: AI voice bots replaced 38% of manual collection calls; QA audit coverage scaled 12% → 100%
• Built a PyTorch signature-verification model and a GenAI portfolio analyzer; delivered predictive (Prophecy) and operational (MOTIF) ML platforms