Headline

Data Scientist @ Kotak | Agentic AI & LLMs at Scale | RL | MTech CSA @ IISc | Ex-TCS

About

Data Scientist and ICCV 2025–published researcher specializing in production ML, Agentic AI, and Applied Reinforcement Learning. I bridge state-of-the-art architectures with enterprise-grade financial systems at scale. Agentic AI & LLMs @ Kotak : Building agentic system for credit intelligence and PL underwriting. It is LangChain based with planning, dynamic tool selection and dual-model on-device LLM orchestration (Mistral + DeepSeek via Ollama). Framed out guardrails, Pydantic inter-agent contracts, and trace-analysis-driven debugging to ensure auditibility and reliability of autonomous workflows—improving downstream throughput by ~400x. FinTech at Scale: Built ML systems supporting 60M+ customers and ₹180Cr+ in monthly credit disbursals, with income modeling (SBERT), RAG-based transaction tagging (1.1B+), and unstructured address intelligence (IndicBERT/LoRA fine-tuned)—driving a 40% reduction in portfolio bad rate while owning the end-to-end MLOps cycle (AWS, Airflow, Redshift). Research (IISc Bengaluru): ICCV 2025 work introduced a Fourier Neural Operator encoder for model-free visual RL (PPO, A2C, Rainbow), achieving SOTA on CARLA and ATARI. It spans representation learning and sample efficiency for RL Core skills: LLMs · RAG · LangChain · Agentic AI · Transformer Fine-Tuning · LoRA/QLoRA · NER · MLOps · AWS · Apache Airflow · PyTorch · TensorFlow · Hugging Face

Location

Bengaluru

AI stack

claude code

claude

cursor

hermes agent

hugging face

ollama

Work experience

Data Scientist at Kotak Mahindra Bank, Jul 2024 - Present · 2 yrs 2 mos, Bengaluru, Building at the intersection of Agentic AI, financial intelligence, and large-scale NLP—autonomous credit-decisioning agents, risk modeling, RAG systems, and production MLOps. - Key Contributions: Agentic Credit Intelligence: Built a LangChain-based agent for natural-language credit decisioning over transaction and CIBIL data. Goal-conditioned pipeline with dynamic selection across 25+ analytics tools and dual on-device LLM orchestration (Mistral + DeepSeek via Ollama). Added guardrails, Pydantic inter-agent contracts, and trace-analysis for observability—delivering autonomous reports (HTML) and boosting Hindsighting throughput ~400x/day. Income & Affluence Intelligence: Replaced heuristics with an SBERT-based (Siamese-BERT) salary detection stack—lifting accuracy 47%→72% and cutting bad rate 1.24%→0.75%, powering ₹180Cr+ monthly disbursals across CC, PL, and BL portfolios. RAG Transaction Tagging: Built a RAG engine categorizing 1.1B+ transactions into 25 classes for 60M+ customers, replacing legacy regex and cutting class confusion 23%. Multi-modal encoder (BERT+MLP, late fusion) with FAISS retrieval and few-shot prompting (Qwen), raising coverage for complex classes 45%. Address Intelligence: Fine-tuned IndicBERT via LoRA for NER (82% token accuracy across dialects) and built a geo-spatial layer bucketing applicants into 0/1/10km zones—raising correct rejections 34% to 64% and cutting wrongful rejections 20% to 7%. MLOps at Scale: Architected AWS Redshift + Airflow pipelines with automated schema validation and data-shift detection, cutting runtime from 5 days to <8 hours (93% gain).

Assistant System Engineer at Tata Consultancy Services, May 2021 - Jul 2022 · 1 yr 3 mos, Pune, Maharashtra, India

Education

Master of Technology - MTech, Computer Science, Indian Institute of Science (IISc)

Bachelor of Technology - BTech, Computer Science, Mahatma Jyotiba Phule Rohilkhand University