// data science student · genai & agentic systems

I build systems that retrieve, reason, and self-correct.

B.Tech CSE (Data Science) student at Amity University. GenAI intern at NHPC Limited. I build production-shaped AI systems — RAG pipelines, autonomous agents, and deep learning models — and I test every claim I make about them.

Get in touch GitHub ↗ LinkedIn ↗
RAG Studio — retrieval pipeline built during NHPC internship
Scanned PDF
→
Self-Healing OCR
→
Dual-Path Embedding
→
ChromaDB Retrieval
→
Cited Answer
01 — About

Grounded in real systems, not tutorials

I care about the difference between a system that runs and one that's actually correct. Every project on this page has a real debugging story behind it — a bug found, diagnosed, and fixed, with the before/after numbers to prove it. I'd rather show a 72%→100% accuracy improvement with a documented root cause than an unverified "100% accurate" claim.

02 — Experience

NHPC Limited

Gen AI Intern
May 2026 – Jul 2026
IT Department · Navratna PSU · Faridabad, India · 8-week NTCC Internship
  • Designed and built RAG Studio, a production-ready multimodal Retrieval-Augmented Generation system for enterprise document intelligence over PDFs containing text, diagrams, and photographs.
  • Engineered a three-pass self-healing text correction pipeline (regex → SymSpell → contextual LLM repair) to eliminate OCR noise before vectorisation.
  • Built a dual-path cross-modal retrieval architecture using BAAI/bge-large-en-v1.5 (text) and CLIP ViT-L/14 (images) over a ChromaDB vector database.
  • Delivered a fully containerised app via Docker, with a FastAPI async backend exposing SSE-streamed responses with page-level citations.
03 — Projects

Built, tested, and debugged end to end

QueryMind
An autonomous agent that translates natural-language questions into SQL against a 4-table e-commerce database, with a self-correcting plan → execute → verify → retry loop. Built a ground-truth verification framework that distinguishes SQL execution success from true answer correctness — found and fixed 3 distinct silent-failure modes.
72% → 100% accuracy
Python · Groq · SQLite
View on GitHub ↗
Dermatological Lesion Classifier
Compared a custom CNN against a MobileNetV2 transfer-learning model on HAM10000 (10,015 images, 7 classes). Found the custom CNN had higher melanoma recall despite lower overall accuracy — a clinically meaningful tradeoff. Diagnosed and fixed a double-normalization bug that had been crushing model accuracy to near-zero.
70.5% acc · transfer
TensorFlow · sklearn
View on GitHub ↗
04 — Skills & Certifications

Tools I can actually defend in an interview

Languages
PythonSQL
AI / ML
TensorFlow/Kerasscikit-learnRAG SystemsAgentic AITransfer Learning
Tools
FastAPIDockerChromaDBGroq APIStreamlit
Cloud & Networking
AWSCCNA FundamentalsTCP/IP
AWS Certified Cloud Practitioner
Amazon Web Services · 2026
CCNA: Introduction to Networks
Cisco Networking Academy · 2025
Microsoft Certified: SQL AI Developer Associate
Microsoft · 2026