Senior Data Scientist · Kochi, India
Seven years building and shipping AI systems in healthcare, industrial automation and e-commerce. I own delivery end to end: requirements, architecture, modelling, deployment and monitoring on AWS.
- Agentic systems — LangGraph supervisor agents, tool calling, LLM-based routing, MCP client integrations, workflow automation for manufacturing and back-office processes.
- Retrieval-augmented generation — grounding LLM responses in live PostgreSQL, queues, APIs and vector stores (FAISS, Qdrant, pgvector), including self-hosted open-weight models to control inference cost.
- Document intelligence — spaCy named-entity extraction, summarisation and Tesseract OCR pipelines over unstructured clinical reports, running as asynchronous, message-driven services.
- Speech and voice — real-time STT → LLM → TTS voice agents with conversation state and interruption handling, including Malayalam and other Indic languages.
- Computer vision — CNN and transfer-learning models for clinical and sensor imagery, served behind Flask and FastAPI.
- Production engineering — containerised services on AWS (Lambda, S3, serverless), RabbitMQ, Docker, CI/CD, and leading cross-functional backend, data and frontend teams.
Most of my production work is proprietary. The repositories below are personal projects that show the same patterns in the open.
| Repository | What it shows | Stack |
|---|---|---|
| queryflow-rag-chat | Streaming RAG chat with hybrid BM25 + dense retrieval, abstention when nothing is relevant, and an evaluation harness (recall@k, MRR, LLM-judged faithfulness) gated in CI | FastAPI, SSE, Next.js, sentence-transformers, Docker Compose |
| smarthire-agent-workflows | Config-driven multi-agent workflow engine: YAML-registered agents, tasks and tools; linear, parallel and conditional graphs; tested with fake nodes | LangGraph, CrewAI, Flask, Pydantic |
| voice-guided-surgical-arm | Voice command → LLM structured output → OpenCV tool detection → ESP32 robotic arm; parser evaluated on labelled spoken instructions | LangChain, Ollama/OpenAI, OpenCV, Flask, Arduino |
| disaster-tweet-classification | Five models from TF-IDF to fine-tuned DistilBERT under one evaluation protocol: CV-chosen thresholds, calibration, error analysis showing label noise, served model with per-prediction explanations | scikit-learn, PyTorch, Transformers, FastAPI, Docker |
| multi-persona-local-chat | Persona-switching chat over local DeepSeek models | Streamlit, Ollama |
| Languages and data | Python, SQL, PostgreSQL, MongoDB, FAISS, Qdrant, pgvector |
| Generative AI | LangChain, LangGraph, CrewAI, MCP, OpenAI, Claude, Ollama, Hugging Face, open-weight models |
| ML and NLP | TensorFlow, Keras, scikit-learn, spaCy, CNNs and transfer learning, NER, summarisation, OCR |
| Speech | Sarvam AI, STT/TTS pipelines |
| Backend | FastAPI, Flask, Django, REST, RabbitMQ, asynchronous processing |
| Cloud and MLOps | AWS, Azure, GCP, Docker, CI/CD, model deployment and monitoring |
Microsoft Certified: Azure AI Engineer Associate (AI-102).