Cofounder & Data Analyst · Aug 2023 – Sep 2026

Baby Steps IVF

makebabysteps.com · Portland, OR

Patients choosing an IVF clinic see raw success rates, and those rates reward clinics for treating easier patients. Baby Steps compares clinics on outcomes adjusted for the patients each clinic actually treats.

3M+IVF cycles

400+US clinics

28 yrsof CDC clinic reports

~50pipeline steps, run weekly

🎯What it does

Baby Steps IVF helps people compare US fertility clinics fairly. I designed the data foundation and the statistical models behind it, set product priorities, and coordinated two developers against a weekly release.

🧭How the data flows

SourcesSART registry, CDC reports, Census, CMS, NPPES, FDA, price files

→

Weekly pipelinePython pulls, cleans and joins every source

→

Checks & reviewRe-certify each clinic-year; hold disputed records for a human

→

PublishPostgreSQL (Supabase) and a hosted prediction API

→

SiteClinic comparisons on makebabysteps.com

📊Data

• Over 3 million IVF cycles from 400+ US clinics, from the national SART registry.
• 28 years of CDC clinic reports, harmonized across reporting formats that changed over time.
• Census and community-health covariates, so comparisons aren't driven by differences in the populations clinics serve.
• IVF prices: CPT/HCPCS rates from hospital price-transparency and insurer Transparency-in-Coverage files, matched to clinics.

🔗Record linkage

Clinics rebrand, merge and sometimes register twice. I linked SART and CDC clinic records across rebrands, mergers and duplicate registrations, and logged the evidence behind every match decision so each one can be reviewed.

• Rebrands
• Mergers
• Duplicate registrations
• Evidence log

🧬The IVF funnel, modeled stage by stage

Each cycle is modeled one stage at a time, so the site can show where each clinic's patients drop off, not just how many end in a birth.

Schematic of the modeled stages (not drawn to scale).

💠Modeling and validation

• Risk adjustment: Bayesian hierarchical models adjust each clinic's success rates for patient age and case mix, so clinics treating harder cases aren't penalized.
• Honest validation: models train on 2014–2022 and are scored on 2023–24 data they never saw.
• Gradient-boosted baselines (XGBoost/LightGBM) with walk-forward validation for comparison.

Each year of registry data, by its role in model fitting. Hover a year for details.

🧰Tools

• Python
• pandas
• Polars
• NumPy
• PyMC
• XGBoost
• SQL
• DuckDB
• SQLite
• PostgreSQL (Supabase)
• Modal
• Git/GitHub
• GitHub Actions CI

← Back to portfolio