I build AI systems that hold up outside the notebook: agentic RPA that replaces costly bot licences, MCP tool servers and LLM gateways that let agents act inside real enterprise software, retrieval-driven LLM systems, and computer vision at the edge. I start with the business case (process discovery, scope and ROI), then architect, evaluate and ship.
- 🏢 AI/ML Engineer @ NGXP Technologies (Kochi): leading AI & automation discovery for a large enterprise. 600+ processes assessed and 100+ process cycles shortlisted for rollout, each with scope and ROI.
- 🤖 Designing an agentic RPA platform: custom LLM agents on the open-source Cua computer-use framework, cutting the per-bot licensing cost of tools like UiPath.
- 🔌 Architecting MCP tool servers for ERP/CRM behind an LLM gateway (routing, fallback, cost control) with human-in-the-loop approvals.
- 📏 Defining AI evals + LangSmith tracing to verify agent accuracy before rollout.
| 3+ yrs | building production AI, from edge CV to agentic systems |
| ~80% | lower LLM inference cost via 4-bit quantization + RunPod serverless |
| 95% | accuracy in deployed real-time vision models |
| 180+ FPS | zero-GPU ANPR on CPU (HOG + PaddleOCR) |
| ~58% Acc@0.5 | RefCOCO visual grounding on an 8GB GPU (SpatialVLM) |
| Project | What it is | Stack |
|---|---|---|
| Hivefloor | Local-first desktop office for AI coding agents: real terminals, a boss agent that routes work, shared memory, Docker sandboxing, approval gates | TypeScript · Electron · node-pty |
| SpatialVLM | Grounding head on 4-bit Qwen2-VL-7B for pixel-precise object localization | Qwen2-VL · PyTorch |
| VLM-Guided Anomaly Detection | Zero-label defect detection: CLIP patches + VLM normality priors guiding PatchCore, MVTec AD benchmark | CLIP · PatchCore · OpenVINO |
| SpatialVLM Auto-Annotator | Auto-labels images for detection, segmentation and pose with a VLM, no manual annotation | Qwen2-VL · Python |
| WebRTC Video Calling | P2P browser calls + GStreamer pipeline exposing calls as RTSP / YouTube Live for YOLO vision | WebRTC · Socket.io · GStreamer |
| Transcript → Knowledge Graph | Whisper transcription → semantic chunking → local-LLM extraction → Neo4j graph | Whisper · Ollama · Neo4j |
| Medical NLP Chatbot | Clinical Q&A with hybrid retrieval (SQL + vector RAG) and a local LLM | FastAPI · ChromaDB · SQLite |
Agentic AI & LLMOps: MCP Servers · Agentic RPA (Cua) · Computer-Use Agents · LLM Gateway · AI Evals · LangSmith · LangChain · LangGraph · Code Mode · OpenCode
GenAI & LLMs: Claude API · LLaMA · Qwen2-VL · Hugging Face · Ollama · vLLM · RAG · 4-bit Quantization · NLLB
Computer Vision & Edge: YOLO · OpenCV · DeepStream · OpenVINO · TensorRT · GStreamer · NVIDIA Jetson · Pose Estimation · Knowledge Distillation
Data & Infra: Docker · AWS · RunPod Serverless · FastAPI · Kafka · WebRTC · PostgreSQL · MongoDB · Supabase · FAISS · ChromaDB · Milvus · Qdrant · Neo4j
Built for production, not demos.