About
AI Engineer who builds systems that actually work, translating research into production AI and agentic systems that solve real problems.
At Grant Thornton INDUS, I take messy enterprise workflows and replace them with AI that actually runs in production.
Most of my work involves the engineering side of AI, figuring out how to make LLMs retrieve reliably, reason properly, and use tools without breaking in production. You can find me talking about it on X and LinkedIn.
Experience
Grant Thornton INDUS
AI DevSecOps Engineer
Jul 2026 – Present
- •Implemented a multimodal AI triage pipeline using LangChain and GPT-based models to process email bodies and image, PDF, and Word attachments, automate ServiceNow ticket creation and routing, and reduce estimated processing time from 15–20 minutes to 20–30 seconds per ticket (~98–99% reduction).
- •Built attachment-specific image-analysis and document-extraction workflows, combining email and attachment context with ServiceNow assignment-group knowledge to generate structured, ticket-ready JSON.
- •Integrated prompt-injection screening, sender-domain validation, and automated-email detection into the triage workflow to strengthen security and improve routing reliability.
- •Developed a runbook assistant using Copilot Studio and SharePoint to provide grounded answers from approved procedures through Microsoft Teams, with Power Automate synchronizing updated runbooks to the agent's knowledge repository.
Neurapex AI
AI Research Engineer
Jun 2026 – Jul 2026
- •Built a Python pipeline for the CT-RATE dataset, automating NIfTI volume downloads, Hounsfield Unit (HU) calibration, three-window CT image generation (lung, soft tissue, bone), and VQA dataset preparation.
- •Adapted Google MedGemma 1.5 (4B) for volumetric CT inference using custom slice selection, multi-window preprocessing, and batch inference; identified and corrected a preprocessing issue that could destroy diagnostic HU information.
- •Investigated M3D-LaMed and Hulu-Med 7B for 3D medical vision-language tasks, examining volumetric preprocessing, visual tokenization, spatial pooling, and vision-language integration; implemented inference scripts and experimented with alternative input representations.
- •Built a radiology report evaluation framework supporting METEOR and BERTScore, with per-volume scoring and JSON/CSV input support to evaluate generated reports against reference reports.
Featured Projects
ClaimLens
Jan 2026 – Mar 2026
Built a production-grade RAG pipeline for insurance policy analysis across multi-insurer documents.
Designed for real-world insurance policy analysis — where naive RAG fails and structure-aware retrieval wins.
- •Built a production-grade Retrieval-Augmented Generation (RAG) pipeline for insurance policy analysis across multi-insurer documents.
- •Implemented a deterministic clause splitter parsing legal PDFs into atomic clause chunks across 5 structural formats with canonical clause IDs and duplicate detection.
- •Designed a two-stage retrieval architecture using embeddings + FAISS (Top-40) with cross-encoder reranking (Top-5), achieving Recall@20: 0.93, MRR: 0.89 on single-clause queries.
- •Developed a multi-clause evaluation framework with stage-wise diagnostics, achieving Coverage@20: 0.87, Full Recall@20: 0.60, and MRR: 0.77 on composite multi-hop queries.
- •Developed an LLM reasoning engine with strict Pydantic schema validation, citation grounding, and JSON retry logic returning structured answers with confidence and clause citations.
Insurance-Aware RAG
Jan 2026 – Feb 2026
Advanced RAG system for extracting and reasoning over complex insurance policy clauses.
Improved retrieval reliability in complex legal documents with deterministic, traceable clause parsing.
- •Implemented deterministic clause splitting with stable canonical IDs (e.g., ICICILombard_p8_Grace_Period_1) to ensure perfect evaluation traceability and prevent vector overwrites.
- •Built a two-stage retrieval pipeline: high-recall dense FAISS retrieval using BGE embeddings followed by cross-encoder reranking for precision evidence surfacing.
- •Enforced strict Pydantic schema validation (RAGResponse, Citation) as a structural firewall, rejecting hallucinated citations and logically contradictory LLM outputs.
Skills
LLM / GenAI
Machine Learning
Data & Analysis
Backend
Languages
Achievements
Education
Indian Institute of Technology Patna
Master of Technology in AI and DSE
Jan 2026 – Present
Lovely Professional University
Bachelor of Technology in Computer Science and Engineering
Aug 2020 – Aug 2024