5 Years in Regulated FinTech & AI Engineering

Senior Data Scientist – AI
Production GenAI & LLMs

5 years of experience across TransUnion and Bajaj Finserv Health. "I've written the ETL and I've written the LLM prompts. I've built the API, the database, the dashboard, and the model that consumed all three. That range is deliberate."

View Work Resume GitHub LinkedIn Visual Guides
35%
Consumer Dispute Volume Reduction
~90%
Manual Handling Saved on Non-Actionable Cases
₹10M+
Fraudulent Transactions Prevented
79%
Anomaly Model Accuracy (2.7% Geo Fraud Flagged)
Engineering Approach

Layer-Agnostic Problem Solving

"The pattern I care about most: knowing when a problem actually needs a model, and when a well-placed rule, a better pipeline, or a conversation with the business team solves it faster and cheaper. Plenty of my best work involved talking someone out of the complicated version. When you've owned a problem end-to-end, you stop being precious about which layer the answer lives in."

Selected Work

Production Systems & Case Studies

GenAI & RAG Platform

Consumer Dispute Resolution GenAI Platform

Architected a modular LLM platform for dispute resolution in a regulated credit bureau: hybrid RAG over policy and regulatory corpora, tool-calling agents for multi-step workflows, and PII/compliance guardrails. Built email triage combining intent classifiers with LLM response generation under guardrails (cutting ~10 min/case). Cut dispute volume by 35% and manual handling by ~90%.

Llama 3 Gemini Hybrid RAG Tool Agents LoRA vLLM
OCR & Document Understanding

Handwritten Prescription Document Pipeline

Built an end-to-end document understanding pipeline converting photographed, handwritten prescription images into decision-ready structured data. Applied Vision OCR followed by LLM-based entity extraction & slot filling (OpenAI APIs) to parse medications, dosages, and diagnoses from noisy text, enriched against master data for claims fraud ML models.

Vision OCR LLM Slot-Filling OpenAI API Entity Normalization Python
Open Models & RL

Self-Hosted Llama 3 & Qwen GRPO Post-Training

Self-hosted Llama 3 fine-tuned with LoRA on vLLM alongside Gemini on GCP Vertex AI for data privacy, cost control, and low latency. Independently post-trained Qwen models using GRPO (Group Relative Policy Optimization) to evaluate where reinforcement learning genuinely beats SFT versus where it burns compute.

Llama 3 vLLM Qwen GRPO (RL) LoRA / PEFT PyTorch
ML & Risk Analytics

Syndicate Fraud & Geolocation Risk Engine

Led fraud analytics charter (2 analysts, 6 interns); built risk scoring combining supervised classifiers, DBSCAN clustering, and LLM entity analysis preventing ₹10M+ fraud. Geolocation aggregation flagged 2.7% extra fraud; mined clickstream data with 79% accuracy anomaly detection models for real-time fraud identification.

DBSCAN XGBoost Anomaly Models (79%) Geolocation Clickstream
Data Engineering & APIs

Near Real-Time Data Warehouse & Blocklist API

Set up near-real-time data warehouse in MySQL (Python + Azure ADF + Power BI). Automated Data Mart creation (<1 hr go-live vs 1+ wk) and implemented automated data lineage & metadata scanning for DataLake & Power BI datasets. Engineered Mobile & Email Blocklist API preventing 200+ escalations/month.

MySQL Azure ADF Data Lineage Flask Power BI
Enterprise Automation

Central Financial Reconciliation & MIS Automation

Deployed Central Reconciliation System saving 50+ manual hours weekly across teams, paired with configurable MIS Email Automation reducing deployment time by 98%. Designed and deployed a Node.js application automating Loyalty Card creation and management.

Python Node.js Flask Automation Docker
Personal Projects

Open Source & Tools

RAG & AI

sql-agent

Premium Multi-Agent Vendor RAG System. A high-performance, observable RAG system designed for complex vendor data analysis, built in Python.

Python Multi-Agent RAG
View on GitHub
Finance & Web

FundMatrix

A local-first personal portfolio dashboard for Indian investments. It parses CAMS Consolidated Account Statement (CAS) PDFs and rebuilds mutual-fund lot history with FIFO matching.

JavaScript PDF Parsing Finance
View on GitHub
Data Visualization

college-data

Interactive data visualization and processing tool for analyzing and structuring college metrics and datasets.

HTML JavaScript
View on GitHub
Web Apps

my-life-helpers

A suite of lightweight, frontend-only web applications powered by Firebase, including a Debt Tracker for managing personal debts and credit card bills.

JavaScript Firebase
View on GitHub
Python Scripting

CAMS_mf_parser

CAS - CAMS+KFintech Parser. A utility designed to reliably parse and extract investment data from complex mutual fund CAS statements.

Python Data Extraction
View on GitHub

Archived Projects

View 13+ older hardware, python, and web projects from my archives.

View Archives
Experience

Work Experience (5 Years)

TransUnion

Sr Data Scientist - AI, Global Data and Analytics
Aug 2024 – Present (2 yrs 1 mo)
  • Architected and shipped the consumer dispute resolution LLM platform: hybrid RAG over policy and regulatory corpora, tool-calling agents for multi-step workflows, PII and compliance guardrails. Cut dispute volume 35%, manual handling ~90%.
  • Self-hosted Llama 3 on vLLM alongside Gemini (Vertex AI) for data residency, cost and latency control; fine-tuned with LoRA on curated dispute data for output consistency.
  • Built the eval harness the platform ships against: golden set stratified across dispute types, retrieval and generation graded separately, LLM-as-judge with human review, regression gates blocking deploys.
  • Built LLM email triage and auto-response - intent classification, queue routing and templated draft generation, saving ~10 minutes per case across 5,000+ cases/month.
  • Added LLM-generated explanations for production ML decisions, turning feature attributions into reviewer-facing rationale; post-trained with GRPO so explanations generalise to unseen decisions instead of reproducing training-set phrasing.
  • Working on goal-seeking agents with persistent memory, goal decomposition and sub-goal recovery for deterministic, auditable multi-step resolution, utilising SOAR cognitive architecture.

Bajaj Finserv Health Limited

Data Scientist
Jun 2022 – Jul 2024 (2 yrs 2 mos)
  • Led a 2-person fraud analytics team. Built the transaction and claims risk-scoring engine - supervised classifiers, DBSCAN clustering and LLM entity analysis, surfacing coordinated syndicate activity and preventing 10M+ in fraudulent transactions.
  • Engineered a document understanding pipeline turning handwritten prescription images into structured claim data: OCR plus LLM entity extraction for medications, dosages, diagnoses and prescriber, consumed as features by downstream fraud models.
  • Added geolocation-based fraud detection aggregating user and provider location signals, flagging 2.7% of transactions that rule-based checks missed; real-time anomaly models at 79% accuracy.
  • Built a full-stack Mobile & Email Blocklist framework - API, schema and near-real-time ETL, preventing 50+ escalations weekly.
  • Shifted 20% of users from reimbursement to cashless claims through behavioural analysis, saving 5M+, with 1M+ more from claim-level process improvements.
  • Provided technical guidance and mentorship for 6+ data interns.

Bajaj Finserv Health Limited

Data Engineer Intern
Sep 2021 – May 2022 (9 mos)
  • Automated Data Mart creation, cutting go-live from a week to under an hour; built a near-real-time MySQL warehouse and a central financial reconciliation system saving 50+ manual hours.
Technical Competencies

Skills & Technology Stack

Languages & Core Tools

Python SQL PySpark Pandas NumPy JavaScript (Node.js) C++

Generative AI & LLMs

RAG LangChain Huggingface Transformers LangGraph Agentic AI Llama 3 Gemini Qwen vLLM Ollama Fine-Tuning (LoRA / PEFT) GRPO (RL) Guardrails

ML, Vision & Data Science

Vision OCR LLM Slot-Filling Scikit-Learn PyTorch TensorFlow DBSCAN XGBoost FAISS Chroma

Data Engineering & Cloud

GCP (Vertex AI, BigQuery) Azure Databricks MySQL Flask FastAPI Docker Kubernetes Power BI
Interactive Learning

Visual Learning Guides

Explore interactive, visual deep-dives into machine learning concepts - covering the HuggingFace training ecosystem, GPU memory & OOM errors, LLM fine-tuning strategies (LoRA/QLoRA), agentic workflow patterns, and ML mathematical foundations.

Explore All Visual Guides →
Credentials

Education & Certifications

Education

B.Tech, Electronics & Telecommunication Engineering

Madhav Institute of Technology and Science, Gwalior, India
2018 – 2022

CGPA: 8.88 / 10
IBM Specialization

Generative AI for Data Scientists Specialization

Issued by IBM (Nov 2025) - Advanced specialization in building, fine-tuning, and deploying generative AI pipelines for enterprise data science.

Verify Credential
IBM Professional

IBM AI Developer Professional Certificate

Issued by IBM (Nov 2025) - Comprehensive certification covering deep learning architectures, Python AI frameworks, and production model deployment.

Verify Credential
Microsoft Certification

Azure Data Fundamentals (DP-900)

Issued by Microsoft (Feb 2022) - Core cloud data architectures, relational & non-relational database systems, and Azure analytics services.

Verify Credential
IIT Distinction

Data Science for Engineers (Topper 2%)

Issued by NPTEL / IIT Madras (Apr 2022) - Awarded Topper 2% distinction for exceptional mastery of core data science fundamentals.

Recognition

Honors & Awards

Champ Award

Received Champ Award for outstanding performance in TransUnion.

Finalist - 30 Under 30 & PropelX

Finalist of 30 Under 30 and PropelX, career fast track programs of BFL and BFHL, among more than 1000 participants.

Corporate Excellence Awards

Received various corporate awards such as the Hero, Limelight, and appreciation award for outstanding performance at BFHL.

Winner - RGPV Hackathon 1.0

First place in Statewide & Gwalior Zone hackathon rounds (Oct 2021).

1st Rank - IETE Zonal Proficiency Test

Secured 1st rank in IETE Madhya Pradesh Chapter Zonal Proficiency Test (May 2021).

Blind Coding Winner

2nd position at BITS Pilani APOGEE 2021 Blind Coding competition (Mar 2021).

Merit Scholarship Awardee

Recipient of MukhyaMantri Medhavi Chatra Yojna Scholarship worth my full B.Tech course fees.

Get in Touch

Contact

Always happy to talk about production LLMs, open-weight fine-tuning, fraud analytics, or what breaks when any of it meets real users.

Email Me LinkedIn