I build AI systems that survive contact with production — agentic RAG pipelines with hallucination grading, hybrid CNN-Transformer vision models, and the FastAPI/Node services that put them behind a real endpoint.

Right now: SDE Intern @ Infosys Springboard, shipping REST APIs, AWS infrastructure and CI/CD. Final-year B.E. CSE @ SIMATS (CGPA 8.74/10).

I care about the unglamorous half of ML — evaluation, grounding, reproducibility, and actually shipping.

current:
  building:  Self-Corrective RAG  → live on HF Spaces
  writing:   HCT-CBAM             → IEEE submission
  learning:  RAG at scale · K8s · LLM evaluation
  open_to:   AI/ML Engineer · Full-Stack · OSS

Agentic RAG over technical docs. Grades retrieval relevance, rewrites weak queries, falls back to web search, and verifies groundedness before answering — an explicit anti-hallucination loop, not naive retrieve-and-generate.

LangGraph FastAPI ChromaDB Gemini/Groq Docker

Fine-grained food classification. 81.4% Top-1 · 95.2% Top-5 on Food-101 across 5 seeds at 31.2M params — +18.3pp over ResNet50, +4.6pp over CMT-S. Pre-tokenization CBAM yields a 1.87× transformer amplification effect.

PyTorch ViT CBAM Food-101

A miniature SIEM. Ingests logs, PDFs and raw text through a 7-stage detection pipeline, correlates stateful attack patterns (brute-force, fail→fail→success), clusters with Drain3, masks secrets before they hit the DB, streams live risk over SSE.

Python Drain3 SSE Docker

End-to-end data science platform — AutoML, SHAP explainability, time-series forecasting and generative-AI summaries over any uploaded dataset.

Streamlit Google Cloud SHAP Docker

Four independent solutions to the same problem — Flask+KNN web app, CustomTkinter dashboard, TTS desktop app, MySQL-backed variant — built as a comparison of recognition approaches.

OpenCV dlib KNN Flask MySQL

A full default SMS handler for Android with on-device spam scoring. Implements all four components Android requires to replace the stock messaging app — miss one and the OS won't even list you.

Kotlin MVVM BroadcastReceiver

docs(security): add investigation report for issue #1905 Security investigation contributed to Microsoft's document-conversion library. (open)

Certifications

Recognition

SDE Intern · Oct 2025 – Present

REST APIs with Node.js · Express · MongoDB AWS — EC2, S3, IAM, CloudWatch CI/CD via GitHub Actions

AI & Prompt Engineering Intern · Aug – Sep 2025

+18% task accuracy from prompt redesign −25% token spend via compression 15 reusable prompt templates shipped