I ship across applied ML and full-stack SaaS β from healthcare biosignals to multi-tenant analytics in production. My work spans two worlds: one day it's a CNN reading single-lead physiological data, the next it's per-tenant isolation in an analytics SaaS. The throughline is the same β turn a hard, real-world problem into a working system people rely on.
- π Currently building β ArbFlow, a multi-tenant GA4 analytics SaaS running in production
- π« Researching β sleep-apnea detection from single-lead signals (paper in progress)
- π± Gravitate toward β early-stage teams where shipping is the job
- π¬ Open to β SWE internships Β· 2026 Β· remote-friendly
| Project | What it is |
|---|---|
ArbFlow π’ Live |
Multi-tenant GA4 analytics SaaS β secure per-tenant data isolation and a clean product-metrics dashboard. Deployed and running in production. Β· Next.js FastAPI PostgreSQL Vercel |
Sleep Apnea Detection Research Β· in progress |
1D CNN detecting apnea from single-lead signals, validated with leave-one-patient-out CV. Accuracy is the wrong yardstick when events are rare β a model scores ~91% by mostly predicting normal and misses what matters. Rebuilding the evaluation around recall & PR-AUC over the apnea class is the core of the paper. Β· PyTorch 1D CNN LOPO CV SHAP |
| Self-Healing MLOps Pipeline | Automated training pipeline that catches data drift statistically and retrains/rolls back before degradation reaches users β making reliability a property of the system instead of a fire drill. Β· Python Docker CI/CD |
| [Pre-Eclampsia Risk Model] | Flags high-risk pregnancies early from routine clinical features β surfacing the signal that sits in checkup data before it becomes an emergency. Β· scikit-learn Pandas |
Languages
Frontend
Backend & Data
ML / Ops
No grand process β just a loop I trust:
1. Find what actually matters β 2. Ship the smallest thing that actually works β 3. Put it in front of real users, fast β 4. Measure β fix β repeat
So far, so good.
I like a clean straight six. Same way I like shipping β line it up, make clean contact, send it.

