12 Production ML Models on Auto-Scaling Infrastructure
Masraff automates expense entry and accounting workflows using computer vision and NLP models. Their AI pipeline processes receipt images, extracts transaction data, and classifies expenses—replacing manual data entry for corporate finance teams.
Before Datazone, their AWS bill scaled linearly with user traffic. Idle compute sat running 24/7. Storage costs grew unchecked. No autoscaling meant over-provisioning for peak load.
The Problem
Inference clusters ran 24/7 at fixed capacity. Peak-hour provisioning meant paying for idle compute during off-hours.
Training data, model artifacts, and logs accumulated without lifecycle policies. Storage costs grew faster than user growth.
Every 10% increase in users meant 10% higher AWS spend. No cost leverage as the product scaled.
What Datazone Did
Training Data in Object Storage
Receipt images and labeled datasets stored in S3-compatible object storage with automatic lifecycle rules. Cold data archived after 90 days.
Spot Instance Training Jobs
Model training runs on spot compute clusters—up to 70% cheaper than on-demand. Datazone handles interruptions and checkpointing automatically.
Auto-Scaling Inference Endpoints
Inference clusters scale from 0 to N based on request volume. During off-peak hours (nights, weekends), compute scales to zero. Peak traffic automatically provisions additional nodes.
What Changed
- ×Fixed compute running 24/7
- ×Storage costs growing unchecked
- ×AWS bill scaling linearly with users
- ×Over-provisioned for peak load
- ✓Inference scales to zero during off-hours
- ✓Lifecycle policies archive old data automatically
- ✓4x reduction in overall cloud spend
- ✓Compute matches actual usage patterns
Masraff now runs 12 production ML models on infrastructure that costs 4x less than their previous setup. Auto-scaling handles traffic spikes without manual intervention. Storage costs dropped 10x with lifecycle policies. Their AWS bill no longer scales 1:1 with user growth.
In their words
Datazone became our production home for AI. We deploy our ML models, validate every release with smoke tests, and serve millions of requests daily on top of it. The reliability at that scale is what made us stop worrying about infrastructure.
Onur Sefa Ozcibik
AI R&D Team Lead, Masraff