I close the gap between what respiratory medicine needs and what clinical software delivers — with tools I build myself, and judgement labs and companies can hire.
Most medical AI is built by engineers who can't read a flow-volume loop, and validated by no one at the bedside. I sit on the other side of that line — an ex-clinician in respiratory medicine who builds production tools and can tell you where a model is not just imperfect, but clinically dangerous.
Stress-testing diagnostic and report-generation models against real clinical logic before they ship. I find the failure that statistics hide.
The interpretation engines behind this site, configured for your lab — house style, thresholds and clinical narrative encoded, hours of reporting compressed to seconds.
A first-person account of what a non-coding clinician can build with an LLM — for conferences, CMEs, and teams that want their own doctors building, not waiting.
Each one encodes published evidence and years of clinical decision-making — no proprietary hardware, no hospital IT. Two run live in my own practice today; the rest are interpretation engines available for collaboration and deployment.
The first multilingual acoustic lung-function screening tool. Records voice through any microphone and extracts validated biomarkers — YIN pitch, jitter/shimmer, MFCCs, harmonic-to-noise ratio — to estimate lung function. No spirometer, no clinic, no hardware.
End-to-end polysomnography reporting for diagnostic and split-night titration studies — sleep architecture, stage-wise AHI, oximetry, PLM, arousals and nocturnal BP through to clinical impression and management points. Paired with hypoxic-burden and OSA endotyping calculators.
Cardiopulmonary exercise testing reports built on the Wasserman 9-panel and Seiler training frameworks — generating both clinical interpretation and athlete-facing training-zone reports across age groups from a single dataset.
One pipeline from raw session to signed report: ATS/ERS-2022 PFT interpretation with LLN/Z-score pattern classification and BDR analysis, a real-time spirometry QC gate, and an oscillometry interpreter for small-airway phenotyping.
AI-assisted analysis of sputum microscopy across Petri-dish, hemocytometer and stained modes (Leishman, MGG, Diff-Quik), with an expert-validation module for AI-vs-human differential counts — feeding a branded report pipeline with phenotype-discrepancy flagging.
Interactive navigator for the 2024 ERS/EULAR guidelines on connective tissue disease–associated ILD — covering SSc, RA, IIM, MCTD, SjD and SLE across structured disease tabs.
Longer-horizon work: INSPIRES, building a validated pan-India spirometric reference equation where none exists, with my SpiroQA tool as the pre-analytical quality gate — and LungGym, a pulmonary-rehab tracker with cohort analytics and outcome KPIs.
Most clinicians identify the gap between what medicine needs and what technology provides. Very few close it themselves.
I'm an ex-clinician in Mumbai — a pulmonologist by training, with deep work across pulmonary function testing, oscillometry, sleep-disordered breathing and interstitial lung disease. I never learned to code. I learned to think clearly about clinical problems — and used AI to fill the gap between the problem and the solution.
Each tool was born from a real patient encounter, a real diagnostic question, a real unmet need — and built solo. The scarcest ingredient in medical AI isn't engineering. It's clinical judgement.
For validation studies, report-automation pipelines, speaking, or access to the toolkit — let's talk.