Infrastructure delays were costing Mercury a year of AI progress
Mercury faced a problem common across financial services. It needed modern AI infrastructure while operating in a highly regulated industry.
Its technology stack was built for IT, not for engineers. Compliance requirements made it hard to build anything from scratch, and infrastructure changes through traditional IT workflows took several months.
The AI team had no dedicated platform engineers or infrastructure architects. Every new tool needed its own security review, and the team was stuck.
The team’s job was building models that predict risk and prevent fraud. Learning Kubernetes and managing cloud infrastructure pulled time away from that work. Mercury needed to move at the pace AI tooling moves and still meet enterprise security and compliance standards.
Without Saturn Cloud, I think I may be one year delayed in having whatever we built so far.
AI infrastructure inside Mercury’s own VPC, at startup speed
Mercury deployed Saturn Cloud inside its own virtual private cloud. It started as a development workspace for the AI team and grew into the foundation for Mercury’s AI stack.
Saturn Cloud’s VPC deployment passed Mercury’s infrastructure and security reviews once. After that, any tool or framework built on top counted as compliant, and the team stopped reviewing each addition one by one.
What Mercury runs on Saturn Cloud
Hosted and connectedOne review covered the platform. Everything above it inherited the approval.
Several months of IT workflow became less than a week
Deployment speed went up 10x, and the effect reached well past deployments.
We needed infrastructure that enables us to do AI testing and exploration very, very fast. Things change so quickly.
Onboarding became instant. New team members cloned an existing environment instead of spending up to a month setting up a local machine with libraries and connections.
The AI team doubled in size with Saturn Cloud as its default environment. It tested seven to nine AI frameworks in six months, work that would have taken two years through traditional IT channels.
The platform also ran with zero interruptions. No downtime, no system failures and no internal ops team to keep it running. The team spent its time on risk and fraud models instead of Kubernetes and cloud infrastructure.
Anomaly detection caught auto theft losses hiding in the data
Anomaly detection is one of the workloads Mercury runs on this stack. Next, Mercury plans to extend it to enterprise-wide KPI monitoring.
Fine-tuned models on sensitive data, served on GPUs
Mercury’s roadmap builds on the same foundation.
- Host proprietary fine-tuned models for sensitive company data
- Expand anomaly detection to enterprise-wide KPI monitoring
- Keep evaluating new AI frameworks as the landscape changes
- Add GPU capacity for fine-tuning and inference workloads
The whole stack has evolved so quickly – every month there will be some new open-source tools released, and six months ago’s stack becomes a little bit obsolete. Saturn Cloud’s infrastructure helps us update our stack pretty fast. Very nimble.
We serve open models on dedicated GPUs, from our catalog or your own fine-tuned weights, with throughput and latency written into your contract. When your security team needs it, we deploy inside your own network, including air-gapped.