Forward Deployed Engineer @ Kyndryl · Data Scientist · Agentic AI, GenAI & LLMs
By day I design and deliver enterprise AI for global clients. Here I build the same kind of systems on free infrastructure, so anyone can clone one and run it in two commands — no GPU, no managed services, no monthly bill.
📍 Bengaluru, India
A series of production-shaped AI systems, each taking on an input type the last one couldn't handle. Each repo stands completely alone — no shared runtime, no shared database, nothing to install first. What carries across is the approach, not the code.
| # | Project | What it does | Stack | Status |
|---|---|---|---|---|
| 1 | WarehouseOps AI | Multi-agent warehouse operations — inventory, equipment, safety and forecasting questions answered from live SQL, plus invoice extraction. Bring your own CSV and it answers about your warehouse. | LangGraph · MCP · Groq/Ollama · ChromaDB · SQLite | ✅ Shipped |
| 2 | Video Incident Intelligence | Ask hours of camera footage a question in plain English. You write the checks — safety, stock counts, product types — and every answer cites the frame that proves it. A million frames become a few dozen API calls. | Gemini vision · OpenCV · FastAPI · React · SQLite | ✅ Shipped |
| 3 | Voice operations agent | Hands-free floor assistant for people wearing gloves — local speech-to-text, free TTS, Hinglish. | Whisper · Edge-TTS · FastAPI | 🔨 Next |
| 4 | Catalogue enrichment from a photo | Product photo in, full marketplace listing out — title, description, attributes, English and Hindi. | Vision LLM · structured output | 📋 Queued |
| 5 | Document intelligence + review loop | Extract → validate → route → human review. The reviewer queue that document-AI demos always skip. | OCR · LLM extraction · confidence routing | 📋 Queued |
- Answers cite their evidence. The model writes the sentence; the source is real and you can click it. If it can't show you, it doesn't claim.
- The limitations are in the README. Every repo says plainly what it gets wrong and where it shouldn't be trusted. That costs nothing and it's the part people actually respect.
- Bring your own data. Every project takes your file, not just my demo. A system that only works on the sample isn't a system.
Why free infrastructure? Most enterprise AI architectures assume a GPU cluster, a managed vector database and a paid inference endpoint — which means almost nobody can actually run them. Rebuilding under a hard ₹0 constraint forces the real engineering decisions: what to send, what to skip, what you genuinely need a database for. That constraint is the point.
- Enterprise AI delivery — architect and ship LLM, RAG and agentic solutions from proof-of-concept to production, on-prem and cloud, including regulated industries
- Document intelligence — multilingual pipelines combining OCR, translation and structured extraction
- GPU-efficient inference — model serving with vLLM, quantization, distributed processing
- Forecasting & analytics — 4 years of GenAI + ML demand forecasting for global supply chains: SKU-level ensembles, time-series models, LLM explainability layers
- MLOps — MLflow, CI/CD, automated retraining, model governance at production scale
- Kyndryl — Lead II, Forward Deployed Engineer (2026–present): enterprise AI & GenAI for global clients; technical delivery lead and mentor
- UPL — Data Scientist (2022–2026): end-to-end ownership of GenAI/ML forecasting for supply chain decisions
- HPCL — Data Science Intern: NLP and semantic search over large-scale survey data, extended into my Master's dissertation
- Education — MA/MSc Analytics, Tata Institute of Social Sciences (TISS), Mumbai
- UPL OpenAg Award — for leading the AI-based demand forecasting project
- Databricks GenAI Innovation Award 2025 — part of the UPL team recognized for enterprise LLM adoption
- DeepLearning.AI TensorFlow Developer Specialization
Python LangGraph LangChain CrewAI FastAPI React Azure OpenAI Gemini Groq Ollama vLLM RAG NeMo Guardrails ChromaDB/Qdrant OpenCV MLflow Databricks Prophet scikit-learn TensorFlow SQL Docker GitHub Actions
Twitter Sentiment Analysis · LSTM Stock Prediction · Customer Segmentation · Big Mart Sales · RAG over 10-K filings