I'm a final-year B.Tech CSE (AI/ML) student at SRMIST, Chennai, graduating June 2026. Right now I'm an AI Research Intern at DRDO's DYSL-AI Lab, working on audio-language model evaluation — building bilingual (Hindi/English) benchmark datasets and studying failure modes like hallucination and confabulation in models like SALMONN-7B, Audio Flamingo Next, and Qwen2-Audio.

Outside of that, I build full-stack AI applications end-to-end — from RAG chatbots in production to multi-agent pipelines that audit their own outputs for bias. I care about whether a model's output is actually correct, not just fluent — that's the thread connecting most of what I work on.

• 🔭 At DRDO, built a data pipeline (yt-dlp, ffmpeg, ytmusicapi) that curated a 600-song audio dataset with a 48× speedup over manual collection
• 🧪 Designed a data validation system that auto-corrected 60% of broken records, cutting manual review down to under 8% of the dataset
• 🎧 Benchmarked three audio LLMs (SALMONN-7B, Audio Flamingo Next, Qwen2-Audio) for gender classification and failure analysis
• 🎓 B.Tech CSE (AI/ML), SRMIST — CGPA 8.1

Languages

AI / ML

Backend / Frontend

Cloud / Infra

Certifications: Google Cloud/DeepMind — ADK Agents (Intermediate), SLM Training (Advanced), Intro to Generative AI · AWS Academy — ML Foundations, Cloud Operations

🧠 Memory AI — Persistent RAG Assistant A production-ready RAG chatbot with LLM-powered persistent memory, built on React 18 + FastAPI with the Gemini API and Firestore. Context window management for multi-turn dialogue via cosine similarity search over a vector database of embedded conversation history. AGPL-3.0 licensed. 🔗 Live demo

🔍 LLM Multi-Agent Pipeline for Bias Detection A 4-agent orchestration pipeline (Generator → Auditor → Refiner → Verifier) built with Ollama to detect and fix biased language across 4 demographic categories. Modular agent orchestration with prompt engineering and automated evaluation — surfaced hallucination failure modes in agent-to-agent communication, including the Auditor fabricating verbatim "evidence."

🪸 Coral Reef Classification using ResNeSt + CSA — Accepted at ICCBI 2026 A ResNeSt-50 architecture with a novel Cross-Spatial Attention module for automated coral health classification from underwater imagery. 92.4% accuracy (92.0% precision, 91.7% recall) on 7,384 images, outperforming ResNet-50 (87.2%) and ResNeXt-50 (89.1%); ablation study validated the CSA contribution. Deployed as a real-time inference web app on Flask.

Open to AI/ML fresher roles — always happy to talk about RAG systems, agentic pipelines, or model evaluation.