Recruit Reveal - Low-Latency Model Serving
Machine learning model deployment on Databricks with Python SQL UDFs for real-time NFL draft predictions.
Role: ML Engineer / Data Engineer•2024
XGBoost Multi-class
Model Type
QB/RB/WR position prediction
Reproducible
Feature Engineering
Z-scores, OOF encoding, missing flags
MLflow
Model Tracking
@production alias deployment
SQL UDFs
Serving
JSON probability responses
Technology Stack
DatabricksPySparkSQL WarehouseUnity CatalogPythonXGBoostMLflowNode.jsExpressNext.js
Contents
Problem
NFL draft analysts need real-time player evaluation tools that can predict draft position and success probability across different positions. Traditional scouting relies on subjective evaluation, while data-driven approaches often lack the speed and accessibility needed for live draft analysis.
Architecture
Built an end-to-end ML pipeline on Databricks with real-time serving capabilities:
Raw Data → Feature Engineering → XGBoost Training → MLflow Registry → SQL UDFs → REST API → Next.js UIResults & Impact
Model Performance
- Cross-validation accuracy: 78% across all position predictions
- Position-specific performance: QB (82%), RB (76%), WR (74%)
- Sub-second latency: <200ms response times for prediction requests