About
I am a Medical Doctor and Internal Medicine Resident, but I’ve always been a tinkerer at heart. My journey into tech started before medical school by rooting a Samsung Galaxy S3 without even touching a computer.
From breaking things to understand them, to editing magazines, to acting as the universal tech support wherever I go, I thrive on asking “why.”
Today, I build complex, vibe-coded digital tools using agentic AI, Docker, Linux environments and Google Cloud. I am a fast learner, easily distracted by fascinating new rabbit holes, and driven by the desire to build things that matter — whether in the physical world of medicine or the digital world of software.
My work sits at the intersection of medicine and artificial intelligence: I build clinical decision-support tools, fine-tuned medical language models, and research agents that apply large language models, retrieval-augmented generation, and AI agents to real problems in internal medicine and critical care — always with a deterministic safety layer between the model and the clinician.
Projects
TransBench
Python · LangGraph · MCP / Claude Science
A clinician-to-bench research agent, built as an MCP connector for Claude Science. You give it a bedside observation — say, an antihypertensive that seems to help an unexpected condition — and it returns a verified research brief: mechanistic hypotheses grounded in real, resolvable citations pulled live from PubMed, ClinicalTrials.gov and Europe PMC. An 8-agent LangGraph pipeline runs across Claude Opus 4.8, Sonnet and Haiku, and every hypothesis must clear three rigor gates — entailment, grounding and novelty — before it survives. When evidence is strong it designs a runnable single-cell experiment against a named public dataset; when evidence falls short it refuses rather than hallucinate one.
View source on GitHub
WristPilot
Kotlin · Wear OS · Claude Agent SDK
A Wear OS app that remote-controls Claude Code sessions from your watch. When the AI coding agent on your computer needs something — permission to run a command, an answer to a question — it reaches your wrist as a push notification, and you approve, deny, reply or dictate follow-up instructions by voice. A Node.js bridge drives sessions through the Claude Agent SDK; a self-hostable Firebase relay forwards traffic. Everything is end-to-end encrypted: X25519 pairing verified with a 6-digit code, then AES-256-GCM — the relay only ever sees ciphertext and auto-deletes it within 24 hours.
View source on GitHub
Iatronix
Python · FastAPI · Next.js · pgvector
A clinical reference tool for doctors. You type a clinical question and it fetches live data from FDA, PubMed, PMC, NICE, DailyMed, RxNorm and MedlinePlus in parallel, scores each source for evidence quality, then formats the result with Level of Evidence (LOE I–III) and Class of Recommendation (COR I–IIb) attached to every claim. It also includes a spirometry analysis module that classifies obstructive, restrictive and mixed patterns. A semantic cache (pgvector + Redis) returns repeated similar queries in ~200ms without hitting the LLM.
View source on GitHub·Live site
IatrogeniX
Python · Gemma 4 · LoRA / Unsloth
A fine-tuned medical language model built on Google’s Gemma 4 (2B parameter) architecture, trained on MedQA-USMLE and MedMCQA with LoRA adapters via Unsloth, then quantized to Q5_K_M GGUF for edge deployment. On top of the model sits a deterministic SafetyValidator that checks every output against a drug reference database (1,000+ entries) and ICU protocol rules before it reaches the user. Submitted to the Kaggle “Gemma 4 Good” hackathon; benchmarked against 20+ models and ranked #1 in the sub-10B parameter class with 83.1% accuracy on MedQA.
View source on GitHub
Aegis ETL
Python · FastAPI · Presidio · Ollama
An on-premise document ingestion pipeline. You drop files (PDF, DOCX, XLSX, CSV, MSG, EML, scanned PDFs via OCR) into a POST endpoint and it extracts text, detects and masks 20+ PII entity types using Presidio + spaCy, encrypts the original PII with Fernet before it ever reaches the database, generates 768-dimension embeddings via a locally-running Ollama model, and stores everything in PostgreSQL with a pgvector HNSW index. Search uses a hybrid RRF approach combining vector similarity, trigram distance and BM25 full-text scoring.
View source on GitHub
AbbrScan
JavaScript · Chrome Extension
A Chrome extension that detects medical abbreviations on research papers and shows their expansions in a sidebar. It runs automatically on PubMed, PMC and NEJM pages and does nothing on other sites. Detection uses three methods — dedicated abbreviation tables, inline “Full Term (ABBR)” patterns, and reverse “(ABBR) Full Term” patterns — each scored by confidence. Handles mixed-case forms like mRNA, IgG, siRNA and kDa, and can be toggled per site.
View source on GitHub
This Portfolio
React · Vite · Three.js · GSAP
The site this profile belongs to. A React SPA with two themes — a dark “terminal” mode and a light “clinical” mode that follows system preference but can be toggled. Smooth scrolling is handled by Lenis; animations and scroll-triggered effects use GSAP and ScrollTrigger. The “Rabbit Hole” section is a Three.js force-directed graph built from actual YouTube watch-history data, and the Contact section hides a fully working Tetris easter egg.
View source on GitHub·Live site
Experience
Internal Medicine / MRCP Candidate
Park Hospital, Gurgaon, India
2022 – Present
Managing complex internal medicine cases, high-acuity OPD, and inpatient medicine in a high-volume environment.
Medical Officer
Thane Civil Hospital, Thane, India
2021 – 2025
General Medicine and NCD OPD, managed COVID ICU workflows, and participated in various health camps.
COVID ICU Medical Officer
Bhiwandi COVID ICU, Bhiwandi, India
2021
Managed patients on ventilators, intubations and ABG analysis during the peak of the pandemic.
Medical Internship
Rajiv Gandhi Medical College & CSMH, Thane, India
2019 – 2020
Rotated through all clinical departments. Received the Best Medicine Intern Award (2020).