AI/ML Engineer & Software Developer · MS in Computer & Information Sciences
I build software end to end — the model and the product around it. That has meant multi-agent LLM platforms with proper evaluation pipelines, computer vision quantized down to run on a Raspberry Pi, backend APIs handling async orchestration and streaming, and an open-source Python library. What drives me isn't just writing the model; it's engineering the whole system so it's robust, scalable, and inexpensive to run. This isn't a career I fell into — it's the thing I'd be doing anyway.
- AI/ML across the whole spectrum — classical ML (recommendation systems, similarity scoring), deep learning (YOLO object detection, LSTM/GRU language models), and modern LLM systems (multi-agent orchestration with LangGraph/LangChain, RAG pipelines, embeddings, vector stores). I've built RAG five different ways from scratch just to understand the trade-offs.
- Full-stack engineering around the model — backends with FastAPI and async workflows, REST APIs, OAuth 2.0 integrations, SQL and NoSQL data layers, dashboards, and Docker/Kubernetes for deployment. I also maintain shellxec, an open-source Python library for shell command execution.
- Engineering under constraints — I quantized a YOLOv8 model (INT8/INT4) to run on a Raspberry Pi 4: ~75% smaller, ~95% mAP retained, sub-second inference. Making things work on a $35 computer teaches you more than any A100 will.
- Measuring what I build — LangSmith tracing and Ragas metrics (faithfulness, answer relevance, context precision) baked into my LLM projects. If it isn't measured, it isn't done.
- Wandering into new problem spaces — NLP for Marathi and Hindi, financial research agents, IoT retail systems, resume parsing and job recommendation engines. The unfamiliar domains are where the fun is.
| Project | What it does |
|---|---|
| StockAnalyst-AI | Multi-agent LangGraph research platform analyzing news, transcripts, financials, and macro data; FastAPI + ChromaDB backend with LangSmith observability and Ragas evaluation |
| RAG | Five retrieval-augmented generation architectures built from scratch in Python — no frameworks — tested against real SEC 10-K filings |
| Traffic Sign Classification | Custom YOLOv5 detector for real-time video — 93% classification accuracy, 90%+ mAP — for autonomous-driving use cases |
| QwInSight | PDF document-analysis platform powered by Qwen models |
| shellxec | Open-source Python library that simplifies shell command execution |
| Movie Recommender | Recommendation engine combining collaborative and content-based filtering |
Also building: Butlr, an agentic assistant that monitors Gmail and Calendar, classifies what matters, and drafts replies (FAISS RAG, multi-provider LLM routing) · Smart Cart, an IoT point-of-sale system where the cart detects products as you add them (quantized YOLOv8 on Raspberry Pi, ~95% detection across 50+ SKUs).
ML/AI: PyTorch · TensorFlow · Hugging Face · LangChain/LangGraph · scikit-learn · OpenCV
Backend & data: Python · FastAPI · PostgreSQL · MongoDB · ChromaDB · FAISS
Infra & practices: AWS · Azure · GCP · Docker · Kubernetes · CI/CD · Git/GitHub · Linux · Raspberry Pi
Always happy to talk about ML systems, agents, or a project you're building — LinkedIn is the fastest way to reach me.
The fun is always in doing something that is challenging and tough to do.