AI Overhaul Cuts Costs by 96% and Boosts Accuracy for Identity Verification Platform
47%
Increase in verification accuracy
96%+
Reduction in infrastructure costs
3100%
Platform throughput increase
THE CHALLENGE
When cost, speed and accuracy break at scale
The client’s current AI model was proving to be too slow, too expensive and not sufficiently accurate to meet real-world demand. Their infrastructure and ML workflows were holding back product performance, draining resources and putting the entire business at risk.
Key challenges included:
‒ 2+ days to retrain the model, slowing innovation and iteration
‒ ~50% model accuracy, leading to high manual verification volumes
‒ Slow model runtime (~1 minute per check), damaging user experience
‒ Manual handling of training data, limiting how quickly models could be improved
‒ Cloud costs exceeding $1.2 million/month, threatening financial viability
THE SOLUTION
AI performance, re-architected with MLOps
Catapult introduced robust MLOps practices, optimising infrastructure and redesigning the client’s training pipeline, to unlock massive efficiency gains in both speed and accuracy.
Key solution elements included:
‒ Optimising and tuning the infrastructure by introducing right-sized compute, reduced node usage and transitioning away from GPU-heavy to CPU-based processing.
‒ Moving 100+TB of image data for training the models through building automated data pipelines.
‒ Introducing image classification models, accelerating the sorting and scanning of training inputs.
‒ Implementing real-time feedback loops to enable users to fix poor image quality instantly, during verification.
‒ Optimising preprocessing, layout detection and OCR models for improved speed and precision at every stage.
THE RESULTS
Cutting hours, costs and uncertainty at scale
This wasn’t just a performance fix – it was a business-saving transformation. Hours became minutes, costs dropped by millions and accuracy soared.
Key outcomes:
‒ Verification accuracy increased from 50% to 97%, cutting manual fallout and saving £80k+/month
‒ Model training time reduced from 2 days to 1.5 hours, enabling daily iteration and faster optimisation
‒ 1000× increase in training data improved model accuracy, robustness and performance
‒ Runtime dropped from ~1 minute to ~5 seconds, dramatically enhancing the customer experience
‒ Infrastructure costs reduced by over 96%, from $1.2M/month to just $45k
‒ Platform throughput (TPS) increased by 3100%, supporting massive customer growth without instability
