Data Scientist | AI Engineer · Available for to hire
I'm Elias — a data scientist who turns noisy, half-broken datasets into models people actually trust and decisions teams actually ship. Rigorous where it counts, pragmatic everywhere else.
Nine years turning ambiguous questions into measurable answers. I've shipped fraud detection that cut losses by double digits, forecasting that finance actually plans against, and experiment platforms that keep product teams honest. My north star: work that is reproducible, explainable, and boring in production.
Built an agentic triage system that keeps Facebook's network at capacity — predicting hardware failures and rerouting traffic before a ticket ever fires. Cut monthly incidents 40% and held 98% uptime as the network grew.
Built a family of models that price expected CPM across the US cable-TV market — one per sub-region, because networks set ad inventory locally. Cleaned petabytes of messy impression data for 10k+ channels; the first integration drove $2M in upsells.
Built the pricing model behind a US auto book — turning driver, vehicle and location signals into risk-adjusted premiums. Blended an interpretable GLM baseline with gradient-boosted lift to protect loss ratio while keeping quotes competitive.
Built a tail-risk detector for specialty-line insurance, where a handful of rare catastrophic losses decide whether the book is profitable. Extreme-value modeling plus classifiers tuned for the severe-but-rare class give underwriters an early-warning score.
Built a computer-vision system that assesses car damage from claim photos to decide write-off status — and the salvage value, since the insurer owns the car once the claim is paid. YOLO segmentation trained on a self-built gold-standard dataset paired with historical claims.
Built a calibrated propensity-to-buy model for auto insurance, scoring every prospect's likelihood to convert. Let marketing concentrate spend and quoting effort on high-intent leads instead of the whole funnel.
Built the recommendation engine for Minim, a streaming platform for the smaller artists Spotify couldn't monetize well. Lightweight, user-data-weighted models with online learning cross-pollinated fanbases to drive discovery and keep listeners engaged.
Built a calibrated propensity-to-buy model for health insurance using 2019's state-of-the-art methods, ranking prospects by likelihood to enroll. Made a costly, heavily regulated acquisition channel measurably more efficient.
Built a rare-disease screening model over health data, where positives are a needle in a haystack of hundreds of thousands of members. Imbalance-aware methods surface a ranked shortlist of at-risk patients for clinical follow-up.
Built a health-claims fraud detection system combining supervised scoring with provider-network analysis. Gives investigators a prioritized queue of the claims and collusion clusters most worth their time.
If you have a question buried in your data — or a model that works in the notebook but not in the world — I'd love to hear about it.