Six years across electronics, IoT, Android, data and AI research, and three US AI startups in healthcare, SaaS and EdTech. I ship LLM features end to end: summaries, recommendations, agents, and the data pipelines underneath.
Evaluating and choosing models stage by stage, prompt design with domain experts, and keeping the model off the path when a cheaper step is enough.
Getting reliable structure out of PDFs, scans, images and audio, and retrieval that finds the right chunk, not just a similar one.
Pipelines, standard codes and validation, so that one wrong value never quietly reaches a clinician, an analyst or a customer.
From healthcare AI to analytics agents to teaching. Filter by what you care about.
Contributed to Saya, Nephrolytics' clinical AI: smart summaries of patient history, smart recommendations, model upgrades, and the MCP pipeline behind them.
A Go service mapping free-text lab names from EMR feeds and scans to 59K+ LOINC codes. A cache and skip patterns keep most records off the LLM path.
Compared models for document classification and summarization to replace a single model, giving each pipeline stage its best performer.
Audited every unit and reference range to set per-code bounds, so an impossible OCR value goes to clinician review instead of into a risk score.
LLM agents with LangGraph, Gemini and BigQuery that let analysts query several data sources in natural language, replacing hand-written SQL for routine analysis.
Semantic chunking for RAG over financial reports: sentence embeddings, similarity clustering and token-aware sizing instead of fixed-length splits.
Led a team of 8 developers to ship a notes-generation platform for documents and audio (Whisper, OpenAI), with prompts tailored to students, teachers and admins.
Taught a 40-hour AI/ML curriculum, from CNNs and GANs to LSTMs, to more than 100 students over a summer.
NLP sentiment models, SQL pipelines and Tableau dashboards over 250K+ customer support interactions.
A camera reads the number plate and an RFID line detects a red-light crossing, so the fine is issued instantly. The idea came from standing at a signal in India.
Fans register once; at the gate their face is matched to their ticket, so entry takes seconds instead of a queue.
computer vision · Python · code coming to GitHubA chatbot that answers routine health questions from a patient's records, so a quick question doesn't need another appointment.
chatbot · healthcare · built on demo dataSurvival modelling and statistical analysis to understand which factors shape patient outcomes.
ML · statisticsAWS data pipelines and dashboards, reworked for faster and cheaper reporting.
AWS · ETL · dashboardsOne place for students and skilled workers to compare study and work routes across countries.
prototype in progressA guide for school students curious about AI: what to learn, in what order, and where it can take them.
prototype in progressMedical records arrive as scans, faxes and free text, and every lab writes the same test differently. My job is turning that into clean, coded data a doctor can rely on, and flagging what doesn't add up.
Type a test name the way a report might, and watch it resolve. It runs entirely in your browser.
I have always liked building things. The tools changed: syringes and magnets, then circuits and sensors, then code, then models.
“I am good at the unglamorous middle of a system: the part where data is wrong in ways nobody documented, and being careful matters more than being clever.”
Conferences where builders compare notes on payments, agents and the protocols that connect AI to real systems.
I offer free 30-minute calls for anyone breaking into AI, growing in it, or moving to a new country for higher education. I have been the person figuring it out alone; you don't have to.
Healthcare or not, I would love to hear about it. I am open to new roles and conversations.