AI Tech Lead

Led LLM reasoning distillation with DPO and GRPO optimization using Unsloth and Qwen3, achieving a 78% improvement on GSM8K mathematical reasoning benchmarks through a custom DSPy evaluation framework. Fine-tuned Gemma-4 for enterprise office-document generation using QLoRA and GRPO, reducing inference latency by 55% and cloud costs by 65% across more than 200,000 monthly automation requests.

Architected security, safety, and multi-tenant infrastructure for the Concierto multi-agent platform, including prompt-injection detection, MCP isolation, policy enforcement, and extensible guardrails. Led an enterprise GraphRAG coding assistant and dynamic MCP server platform, improving code retrieval accuracy, lowering developer resolution time, and reducing token and context overhead across 50+ production enterprise deployments.

Architected LLM serving infrastructure on GCP Vertex AI using vLLM and TensorRT-LLM, deploying 13B-parameter models at sub-200ms latency for over 500,000 requests per day. Delivered a 40% cost reduction compared with hosted APIs while maintaining 99.8% uptime over six months.

Orchestrated distributed training with DeepSpeed and FSDP on AWS SageMaker GPU clusters, reducing training time by 65% and scaling to 34B+ parameter models. Modernized the ML platform with Kubernetes, MLflow, and Terraform, reducing model deployment cycles from three weeks to two days and enabling significantly faster experimentation.

Architected multi-agent financial-research systems using LangGraph and CrewAI, automating more than eight data sources and processing over 200,000 tasks monthly with 92% task completion accuracy. Built hybrid financial RAG solutions combining BM25 sparse retrieval and dense vector search for low-latency, high-accuracy answers across more than 500,000 financial documents.

Fine-tuned Qwen3 8B models on domain-specific financial corpora using QLoRA, NF4 quantization, and iterative RLHF, substantially improving entity extraction and reducing hallucinated figures. Deployed models through ONNX and TensorRT INT8 quantization, achieving 3.2x throughput improvement, sub-90ms p99 latency, and more than 500,000 daily inferences at lower cost.

Developed and deployed classical machine learning models in Python using scikit-learn, XGBoost, and Random Forest for classification and regression applications. Built RNN and LSTM NLP models in PyTorch, including preprocessing, embedding, training, and evaluation pipelines.

Engineered full-stack ML applications with React, TypeScript, Node.js, Python, and REST APIs to expose real-time predictions in user workflows. Containerized microservices with Docker and established CI/CD pipelines with automated testing and zero-downtime deployments across staging and production.

Bhopal, India. CGPA: 8.8/10. Also completed an IIT Madras Elite certification in Programming, Data Structures & Algorithms.