Give me noisy data and a vague business problem, I'll find the signal and architect the system that acts on it. 6 years of building production grade intelligent systems across finance, retail, and healthcare, built for the places where trust is non-negotiable: Fraud, Risk, Underwriting & Governance at American Express, Workflow Automation at Jitterbit, Genomic Cancer Detection back at IIT Kharagpur.
From research to production—building scalable AI that delivers.
Production-grade AI systems and research experiments.
Built multi-agent architecture: Orchestrator, NLU, RAG Planner, Executor. Pinecone-backed semantic retrieval improved context precision by over 60% and cut retrieval latency by 30%. Resumable 50+ node workflows with checkpointing.
Enterprise Responsible AI ecosystem with 20+ quantitative and 30+ qualitative evaluations covering hallucination, bias, toxicity, PII, and faithfulness. Integrated Guardrails for automated model gating, reducing governance time by ~60%.
Automated complete API lifecycle management via natural language. Token-aware summarization & memory reduced LLM cost by 60%. Provider-agnostic adapter (OpenAI ↔ Llama ↔ Bedrock) with Langfuse tracing.
LLM-driven news-monitoring pipeline detecting financial distress signals for obligors. Automated early warning system with downstream alerts for credit teams at scale.
Insights on AI engineering, ML systems, and production deployment.
A research engine for the most expensive question in software: build it, or buy it? It does the homework, cites every source, argues with itself — and hands the call back to you.
LLM latency isn't one number — it's prefill and decode, pulling against each other. A mental model for finding the milliseconds before you reach for a quantization library.
Most vector databases run the same core algorithm, so 'which is fastest' rarely decides anything. The differences that bite are operational: who runs it, how it filters, what it costs at scale.
Fine-tuning is the most over-reached-for tool in the LLM toolbox. A grounded guide to deciding between prompting, retrieval, and training — before you spend the weeks.
A journey through innovation, research, and engineering excellence.
Jitterbit
Building iPaaS Bot (multi-agent workflow orchestration with hybrid RAG and checkpointed execution) and APIM Bot (NL-driven API lifecycle management). Provider-agnostic LLM orchestration cut cost ~60% while serving 10K+ daily enterprise interactions, with DeepEval guardrails and Langfuse tracing on CI/CD-backed Kubernetes microservices.
American Express
Led a 5-member team building an enterprise GenAI governance framework (20+ quantitative, 30+ qualitative evals; ~60% faster governance). Delivered NLP solutions for risk monitoring, bank-statement underwriting (~85% accuracy), and complaint intelligence. Designed Transformer-based fraud embeddings outperforming legacy RNNs by ~40 bps.
Algonomy
Fine-tuned BERT and GPT-2 for personalized query auto-completion on a corpus of 1.4M+ retail queries.
National Taiwan University
Big-data genomic analytics on axolotl genome at the Epigenetics Lab. Identified regenerative epigenomic factors through computational analysis.
IIT Kharagpur
CGPA: 7.64/10. Thesis: non-invasive detection of human cancers from multi-genomic TCGA data using deep learning.