Multi-agent architectures, hybrid RAG, and models trained from scratch — grounded in deep-learning research.
More about my research →Five of 30+ — every one open source, every one running in the real world.
Topic in, narrated 1080p lesson out in 2–5 minutes: LLM script generation → Manim animation → Piper TTS → FFmpeg compositing. No cloud, no API keys — runs fully local with GPU acceleration.
Any lecture becomes an interactive study session — synced transcript, semantic chapters, streaming RAG tutor with timestamp citations, auto flashcards. 100% local on a 4GB GTX 1650.
Containerised microservices (Interface / Engine / Voice) for fully private inference — one-command Ollama model integration, low-latency voice, GPU-optimised across Linux, WSL2 and macOS.
Two-stage agentic rewrite (Strategist → Execution) injected into ChatGPT, Claude, Gemini & Perplexity. Multi-provider dispatch in an isolated worker — keys local, zero middleware.
Fail-proof booking engine — atomic transactions and unique constraints verified under concurrent stress tests; JWT auth, rate limiting, resilient React UI, multi-stage Docker build.
30+ production AI systems: hybrid RAG (semantic + FTS + RRF) over Milvus, pgvector, Neo4j and Elasticsearch; LangGraph multi-agent pipelines; 3× inference speedups via layer-wise loading + quantisation; 100% defect detection (40/40) with Laws' Texture Filters.
Leading LLM/SLM research, agentic-AI curriculum and hackathon mentorship for 100+ students. Built Karunya-SLM from scratch — GPT-2 architecture, custom BPE tokenizer, PyTorch training loop, WandB, Open WebUI.
Memory-efficient 70B+ inference, 4/8-bit quantisation, macOS/MLX. 10K+ downloads.
Physics-guided super-resolution for thermal IR: RMSE 0.088K · PSNR 41.44dB · SSIM 0.9775.
Contributor — few-shot LLM benchmarking across tasks (EleutherAI fork).
93.3% efficiency in agricultural AI challenge; 100% defect detection in vision QA.
I'm a B.Tech CSE (AI/ML) student at Karunya Institute of Technology and Sciences, Coimbatore — building AI systems with a bias for local-first, privacy-preserving design: if it can run on your machine, it should.
From training language models from scratch to squeezing 70B-parameter inference onto a 4GB GPU, I work across the whole stack — research, systems and product.
Let time tell.
