ML Platform at Intuit Credit Karma
Rebuilt a 60M-record/day scoring pipeline: 92% faster, 85% cheaper, 10x capacity.
PythonMachine Learning PlatformBig DataGoogle Cloud
Problem
Model inference and scoring on the ML Platform team was taking 7 hours and roughly $600 per run, processing on the order of 60 million records a day. Data science teams wanted to scale their modeling work, but the pipeline was the bottleneck holding them back.
Approach
Modernized the scoring component end to end — removing redundant passes over the data, right-sizing compute to match actual load, and aligning the online and offline scoring paths so experimentation numbers matched production.
Outcome
- 36 minutes instead of 7 hours (92% faster)
- $89 instead of $600 per run (85% cheaper)
- 10x model capacity without 10x cost
- ~$400K/year in projected savings
- Optimizations adopted into Google Cloud's own documentation
Read the full writeup: Scoring 39 million rows in 36 minutes: Optimizing batch inference at Intuit Credit Karma