From Covance/Merck CRA and clinical project management, to writing clinical trial protocols at hospitals, to running my own clinic — now building AI products that solve the problems I lived through. 10 products shipped.
Growing up, my parents told me: if you're not giving it everything, you don't love it enough. The things you truly love, you go all in on.
So whether I was slacking or grinding, they stayed out of it — with one line: you own the consequences.
Over time I got stubborn about it. Once I'm in, I'm in. As a kid I'd spend a whole afternoon on one math or physics olympiad problem — no food, no water, didn't move.
Later, almost by accident, I drifted from the world of math and physics into medicine and medication. The sheer density of knowledge, the discipline of the field — it felt new.
To me, medicine is a loop. From observing the body, to patterns in aggregate data, down into molecular mechanism, back up through how drugs act on the body, and back into the data. That's how humans have learned, all along.
In the AI era, I can turn an idea into a product almost without friction. Commercial or not, building is how I make an idea real.
I love this era. Just like my parents said — when you truly love something, you go all in.
Every product came from a real problem — as a clinician, researcher, or operator. Prototypes marked with ◇ were built for specific enterprise clients.
HIPAA/PIPEDA-compliant EHR. 300 clinicians. Full lifecycle: clinician interviews, engineering coordination, compliance, GTM, ongoing operations. Multi-tenant architecture, data migration, production uptime.
Built for my own clinic. Tracks brand visibility across ChatGPT, Perplexity, and Google AI. SEO alone isn't enough when AI answers are eating search traffic.
11 integrated data sources (FDA Orange Book, PubChem, Espacenet, PTAB, IBM RXN, CourtListener, CMS Spending, SureChEMBL). ML systems: patent type classifier, molecular complexity scoring (RDKit, chemist-calibratable weights), Tanimoto similarity search (Morgan fingerprints), patent challenge win probability model (8-factor weighted scoring), LLM vulnerability scorer, and economic ROI model with 180-day exclusivity scenarios. 13 drugs, 68 patents validated.
Scans your codebase, estimates cloud costs across 25+ services. Published on npm. Integrates with Claude Code and Cursor via MCP protocol.
Enterprise ELN for GMP/GLP environments. Batch experiment records, multi-role access control, regulatory-compliant data export. Replacing paper-based lab workflows.
50%+ of CNS Phase 3 trials fail from endpoint selection. AI platform integrating ClinicalTrials.gov data with LLM analysis for evidence-based endpoint design, site selection, and timeline planning. Companion to MDD-Synth simulation engine.
Monte Carlo trial simulator trained on 97 validated real MDD trials (effect sizes 0.08-1.08). Generates synthetic patient trajectories using Gaussian Copula models (SDV). 4 evaluation tools: trial power simulation, enrichment optimization, sample size calculator, timepoint optimization. Turns endpoint selection from intuition into data.
B2B platform connecting biotech companies with MNC pharma BD executives for drug licensing deals. Real-time buyer interest tracking — which companies view your pipeline, when, and how long. Solving the "black box" in biotech BD.
"Doctors heal. Cheryl handles the paperwork." Chrome Extension side panel for the PA rep at insurance-first telehealth platforms. Paste a clinical handoff, pick drug + payer + scenario, get a payer-aware PA letter back in ~60 seconds at $0.0088 / letter. Built around one hard rule: every sentence has to trace to source — pasted note, payer-policy fragment, trial PMID, or doctor confirmation. No fabrication. A pre-submission compliance gate checks ICD-10 specificity, BMI freshness, contraindication screen, and required sections against each payer's first-pass review before the rep sees the draft. Sections render with per-section confidence (high / medium / low); every rep edit feeds an ML loop that promotes corrections to per-clinic rules after 3 repeats. Measured target is approval rate, not draft quality.
Scope: 6 GLP-1 drugs (Wegovy, Zepbound, Mounjaro, Ozempic, Saxenda, Rybelsus) × 8 payers (UHC, Anthem, Aetna, Cigna, Humana, Kaiser, BCBS, Medicare GLP-1 Bridge) × 10 PA scenarios (initial, renewal, step-therapy exception, denial appeal L1, external IRO appeal, P2P prep, tier / formulary / quantity-limit exception, continuation of therapy).
"Most AI products optimize for sounding right. In medicine, sounding right is the failure mode — it's how a confident wrong answer becomes a denied claim, a missed diagnosis, a harmed patient. So I build the other way around: rules first, model second, every sentence in the output traceable to a source someone can audit. The boring version of safety is the only one that survives contact with a real patient."