Fintech Data Scientist for a new Lending Startup

APPLIED DATA SCIENTIST - PRODUCTION ML FOR SMB LENDING LEAD SCORING Long-term engagement, global remote, async-first. We're looking for a data scientist to own a production intent model end to end. THE WORK We score small-business loan leads for vendor partners who need to know which leads to call first. Our model predicts whether a lead will submit an application for financing, turned into a priority bucket vendors call against - so model quality maps directly to revenue per dial. The production model is a multi-tier union-trained architecture with a rule-based gate that demotes leads that look closed. You'd report to the CTO and own the model end to end - features, training, evaluation, calibration, version handoff, and the contract between training and serving. CONCRETE PROBLEMS ON DAY ONE - Pooled held-out AUC is roughly flat across recent versions. Top-decile capture beats the prior baseline at the very top of the ranking but pooled has plateaued. Lifting this is the headline metric. - Source-class shortcut - structural properties across training sources let any model inflate pooled AUC by learning source identity. Mitigation today is a source-classifier diagnostic plus per-source AUC reporting. The architectural fix is open. - Calibration gap - models rank correctly but probabilities don't match real-world conversion rates by an order of magnitude. Calibrated probabilities unlock downstream product features. - Recoverable enrichment misses - a meaningful share of leads now have match keys that weren't available at original ingest, and re-running idempotent enrichers would recover real coverage. - Form-only tier for unmatched leads - a large share of incoming leads get no useful score because they fail enrichment matching. Coarse firmographic features are available for every raw lead. OUR STACK -Snowflake with dbt for the data layer, medallion architecture -Python end to end - scikit-learn, XGBoost, LightGBM, pandas, numpy. SMOTE and class-weight handlers for imbalance. -Jupyter notebooks for training, production runners for scoring and export -Dagster for orchestration (in build), AWS S3 for inputs and exports EXPERIENCE REQUIREMENTS -4-7 years building production ML systems, including binary classification on imbalanced labels and at least one project where you owned the train/score handoff -Strong intuition for adversarial validation, source shift, and label leakage. If you've shipped a model where pooled AUC and per-source AUC told different stories and you knew which one was real, we want to talk. -SQL and dbt comfort - a meaningful share of feature engineering happens in the dbt layer, not in pandas -Production calibration experience (Platt, isotonic, or temperature scaling) including knowing when each fails -Tree-based modeling depth - XGBoost, LightGBM, RandomForest are our workhorses -Independence in a small team - you'll be the only dedicated DS for now -Plain writing - our docs are direct, our PRs are short NICE TO HAVE -Small-business lending, factoring, or commercial-filing data background -Point-in-time-correct feature engineering experience -Shipped a model where calibrated probabilities were load-bearing (pricing, EV, expected-conversion sizing) -Comfort lightly modifying Python orchestration code (Dagster, Airflow, Prefect) -Owned a model that delivered priority buckets rather than raw scores to an external party LOGISTICS -Global remote, async-first. A few hours of US Eastern overlap required. -Long-term, ongoing engagement. Share your hourly range in your application.