I build real-world AI systems through internships, hackathons, and personal projects, focusing on practical and scalable solutions.
I've worked on:
I'm Oracle-certified in Gen AI and Data Science and love continuously learning and experimenting to build meaningful technology.
Multilingual PII detection and redaction across PDF, TXT, and CSV — layered regex + checksum + spaCy NER + LLM verification, FAISS hybrid-search RAG for document Q&A, a rate-limit-aware model rotation engine, and 120+ unit tests
Hardware-aware hybrid agentic RAG for M&A data rooms — LangGraph orchestration, Qdrant dense + sparse retrieval with cross-encoder reranking, multi-page table stitching with cell-level numeric fidelity, and deterministic hallucination guards; 100% pipeline reliability across validation queries
Safety-first agentic system that extracts and reconciles multi-page hospital PDFs into structured discharge summaries — LangGraph state machine with clinical validation rules; processed a 71-page, 15-document-type record in 208s, flagging 3 genuine conflicts with zero false positives
LLM-driven interview agent with a 7-state machine, 3 concurrent LangChain classifiers per turn (decision engine + sentiment tagger + HR-flag detector), 4 voice modes via WebSocket (STT + TTS), and per-session JSON records, plain-text transcripts, and Markdown HR summaries
7-agent document Q&A system with true hybrid search — 384-dim dense embeddings + BM25 sparse vectors fused via Reciprocal Rank Fusion in Milvus — with semantic chunking, self-corrective query rewriting, and multi-format ingestion (PDF, DOCX, PPTX, Excel)
4-agent marketing system (Analyst → Strategy → Execution → Critic) built with LangGraph + CrewAI — autonomously pressure-tests decisions through a self-corrective critique loop, rejecting poor strategies and replanning up to 3 iterations
Data extraction backbone for the lead generation pipeline — scraping business contact details, emails, and company metadata at scale using Selenium + BeautifulSoup
In-meeting SME bot for Teams/Meet/Zoom that joins as a meeting assistant, listens to system audio, live-transcribes with speaker separation (faster-whisper), detects when it is called upon, and responds with grounded answers in 1–2s using a dual-model draft + refine pipeline with Chroma-based RAG over uploaded documents
AI-qualified lead pipeline scraping Reddit, Discord, Slack, and LinkedIn — with semantic filtering and automated outreach via Twilio, cutting lead generation costs by 94.6%
Confidence & uncertainty-aware dynamic ensemble of InceptionV3, MobileNetV2, and VGG19 for 4-class lung disease classification (COVID-19, Opacity, Pneumonia, Normal) — with temperature-scaled calibration, entropy-based uncertainty estimation, and per-sample adaptive weighting (confidence^α / uncertainty)
Full-duplex real-time voice assistant with barge-in interruption handling — faster-whisper STT, Kokoro TTS, silero-VAD, and multi-provider LLM fallback (Groq, Ollama, Gemini, OpenAI) with exponential backoff, achieving 200–500ms response latency
Built for Indian criminal law assistance, achieving hybrid search across 673 IPC documents with sub-second retrieval
Shamla Tech Solutions
Coimbatore, India
Sri Krishna College of Engineering and Technology
Coimbatore, India
CGPA: 8.0
Oracle · Issued 2025
Oracle · Issued 2025
Udemy · Issued 2024
NPTEL · Issued 2024
Team Lead • Full-Stack Developer • AI Engineer
Led a team of 4 to develop an AI-Powered Crime Detection System using YOLOv8 and CNNs for real-time video analysis and threat identification.
Team Lead • Data Scientist • AI Engineer
Developed AI Legal Engine with RAG architecture achieving semantic search across 673+ legal documents for Indian criminal law assistance.