I work across the full data-to-AI stack and care about the problem more than the title: data engineering, data science, machine learning, and AI engineering. The thread through all of it is the same. I take a real problem end to end, from the data and the model to a working interface, and ship something a person can actually open and use. The work tends to run offline, ground its answers in real sources, and be easy to try in a minute, because that is what makes it worth building. Open to roles at any point on that path, and to any domain.
Calm, disciplined, focused on what I can control.
- Grounded outputs. When a model answers, it should point back at the source it used, not ask you to trust it.
- Local-first by default. Most of my apps run offline with no account and no API key. A single env var opts into a hosted model when you want one.
- Shipped, not notebooks. FastAPI and Next.js, Docker, CI, and in some cases a live URL a non-technical user can open in a browser.
- Across domains. Retrieval, object detection and tracking, knowledge graphs, data extraction, and reproducible dev environments.
| Project | Domain | What it does | Stack |
|---|---|---|---|
| ai-due-diligence-copilot | LLM / RAG | Q&A over deal documents (pitch decks, 10-Ks, term sheets) that cites the source passage for every claim. Runs offline first. | FastAPI, Next.js, Ollama/Claude, embeddings |
| rag-pipeline-langchain | MLOps | Retrieval pipeline on AWS with an MLflow-tracked eval loop and a GitOps deploy path. | LangChain, LangGraph, Pinecone, MLflow, S3/DynamoDB, K8s/Helm |
| federated-yolov8-object-detection | Computer vision | YOLOv8 trained across distributed clients with Flower, so raw data never leaves the client. | YOLOv8, Flower, PyTorch |
| pickleball-vision-llm | Computer vision | Detection and tracking pipeline over match footage, split into modular vision, ML, API, and frontend layers. | Python, OpenCV, PostgreSQL, Docker |
| instagram-reels-extractor | Applied ML / data | Turns reels into structured text (caption, transcript, OCR, vision), renders a PDF and docs site, and adds local semantic search. | Python, transcription, OCR, embeddings |
| pediatric-care-platform | Full-stack app | Local-first desktop app: an AI symptom checker over a symptom-to-disease knowledge graph plus clinical workflows. Synthetic data. | Tauri, FastAPI, Next.js 15 |
| PB Card Deck | Full-stack app | Mobile-first pickleball card game + scorekeeper: draw a twist card mid-match and keep score, no login and no signal needed after first load. Live. | React, TypeScript, PWA |
| bujo | Full-stack app | A private bullet-journal web app with rapid logging and habit tracking. No backend, data stays in the browser. | React, TypeScript, Tailwind |
| rocky-dev-setup | Dev tooling | One command to stand up a Rocky Linux dev environment on mise and chezmoi. | Shell, mise, chezmoi |
More in the repositories tab,
including two more cited RAG services (engineering-intelligence-hub,
healthcare-knowledge-navigator) and loan-division-emi-tracker (variable-rate EMI split
across borrowers with a worksheet view of every calculation).
| Area | Tools |
|---|---|
| Data engineering | PySpark, ETL pipelines, PostgreSQL, DynamoDB, S3, SQL |
| Data science | pandas, NumPy, scikit-learn, exploratory analysis, evaluation metrics |
| LLM / RAG | LangChain, LangGraph, Ollama, Claude, Pinecone, sentence-transformers, citation-grounded retrieval |
| Computer vision | YOLOv8, OpenCV, object detection & tracking, federated learning (Flower), PyTorch |
| MLOps & infra | MLflow, Docker, GitHub Actions, Kubernetes, Helm, mise, chezmoi |
| Backend & apps | Python, FastAPI, Next.js, React, TypeScript, Tauri |
Languages I work with:
- Email: sathishkumar786.ml@gmail.com
- LinkedIn: in/SathishKumarAI
Open to data engineering, data science, ML, and AI engineering roles, and to any domain. If something here is useful to you, reach out.