Headline

Data Scientist @ IQVIA | GenAI & LLM Engineer – Agentic AI, Multi-Agent Systems, RAG, LLM Evaluation | LangGraph · LangChain · MCP · FastAPI · Azure OpenAI | IIT Madras

About

I'm an AI Engineer at IQVIA building production GenAI systems for life sciences and pharma: agentic workflows, RAG pipelines, MCP tooling and LLM evaluation that hold up in front of real users. My working principle: LLMs reason, code computes. Models interpret evidence; deterministic code handles the calculations. Every output has to be grounded, traceable and testable. WHAT I'VE BUILT • Agentic drug-launch forecasting: LangGraph multi-agent workflows combining research agents, validated citations and deterministic forecast models. • Forecast monitoring: MCP-based agents that assess how approvals, trial results and policy changes affect forecast assumptions. • SHAP Assistant (Explainable AI): turns SHAP attributions into plain-English clinical insights, with validator guardrails against unsupported explanations. Cut analysis time from hours to under 50 seconds (~95%), 75% user adoption, 200+ daily queries. • Document intelligence: a custom FastAPI MCP server that turns unstructured PDFs into agent-ready data, enabling autonomous financial-audit workflows with ~90% less manual oversight. • FusionSync: multimodal RAG on a unified FAISS vector store with async orchestration and server-side caching. 60% lower memory overhead, 75% faster downstream processing. • Demand forecasting: statistical, machine learning and deep learning models for pharmaceutical demand planning. HOW I WORK Grounding, evaluation, traceability and failure handling come first, then cost and latency trade-offs. LLM evaluation is my current deep-dive. Outside work, I implemented a GPT-2-style transformer in PyTorch from scratch, including training, text generation and fine-tuning. CORE STACK Python · SQL · LangGraph · LangChain · MCP · Claude Agent SDK · FastAPI · Azure OpenAI · RAG · FAISS · Qdrant · Neo4j · PostgreSQL · Cosmos DB · Docker · PyTorch · scikit-learn · SHAP RECOGNITION • M.Tech, IIT Madras: highest CGPA in my class of 2023 • Gold: IQVIA GenAI Hackathon • B.E., Savitribai Phule Pune University: Department Rank 5 OPEN TO AI Engineer, GenAI / LLM Engineer, Machine Learning Engineer and Applied AI roles focused on agentic systems and LLM infrastructure. 📩 verma.aditya1294@gmail.com GitHub: github.com/jskaditya1294 Portfolio: portfolio-site-tawny-rho.vercel.app

Location

Bengaluru

AI stack

claude code

gemini

github copilot

langsmith

mcp

ChatGPT

Work experience

Consultant - GSK, US at IQVIA, Sep 2026 - Present · 2 mos, Remote

Associate Consultant at IQVIA, May 2026 - Aug 2026 · 4 mos, Bengaluru, SHAP Assistant — Explainable AI Platform (GenAI | FastAPI | Azure OpenAI) Built a production-grade Explainable AI platform that translates complex SHAP feature attributions into plain-English clinical insights — reducing manual analysis time by 95% (from hours to under 50 seconds). Key engineering decisions: • Designed a RAG-inspired pipeline that fuses SHAP outputs with high-dimensional clinical data • Implemented validator-based guardrails to ensure 100% adherence to SHAP-derived evidence, eliminating hallucinations in automated medical insights • Integrated Azure Cosmos DB + Blob Storage for persistent state and explainer caching • Architected a session-aware Dash UI supporting multi-turn reasoning • Achieved 75% user adoption and handles 200+ daily queries at IQVIA Stack: Python · FastAPI · Azure OpenAI · SHAP · Dash · Cosmos DB · Blob Storage · Docker Impact: Solved the "black-box" problem for clinical AI — giving stakeholders trust in automated insights for the first time.

Analyst (Advanced Analytics) at IQVIA, Sep 2024 - Apr 2026 · 1 yr 8 mos, Bengaluru, Context-Aware DI — MCP-Based Financial Auditing Pipeline Architected an MCP (Model Context Protocol) pipeline that transforms unstructured PDFs into queryable, agent-ready assets — enabling autonomous AI agents to perform real-time financial auditing with 90% less manual oversight. Key engineering decisions: • Built a custom MCP Server using FastAPI that bridges enterprise document data with LLMs • Designed for scalability: any unstructured PDF becomes a structured data source accessible to LLM agents • Enabled autonomous validation workflows — agents can audit financial documents end-to-end without human intervention Stack: Python · FastAPI · MCP · LangChain · PDF parsing · LLM agents Impact: Dramatically reduced manual audit effort in enterprise financial workflows — a direct application of emerging agentic AI infrastructure. FusionSync — Multimodal RAG Infrastructure Engineered a multimodal RAG system that synchronises unstructured text and media data into a unified FAISS vector store, purpose-built for agentic workflows. Key engineering decisions: • Built an asynchronous backend orchestration layer with server-side caching to handle heavy embedding maps • Reduced browser memory overhead by 60% through optimised embedding management • Boosted processing efficiency by 75% for downstream agentic workflows Stack: Python · LangChain · FAISS · LangGraph · Async Python · Vector databases Impact: Built the retrieval backbone enabling agents to reason over heterogeneous data sources — text, images, and documents — in a single unified pipeline.

Education

Master of Technology (M.Tech.), Indian Institute of Technology, Madras

Bachelor of Engineering - BE, Civil Engineering, Savitribai Phule Pune University