Case Study: Scaling Surgical Intelligence on AWS for Scalpel AI
Challenge
Scalpel AI has built a computer vision solution that verifies surgical instrument trays before and after surgical procedures. A mobile app captures an image of a surgical tray; inference runs against a suite of machine learning models, and the result comes back in seconds, giving the team confidence that every instrument is present and correctly placed or flagging what isn't.
Errors in tray configuration carry direct clinical risk, and Scalpel's clients, which include healthcare OEMs and large healthcare providers in the United States, depend on the system working reliably, quickly, and at scale. With a critical trial upcoming and more than 30 further clients waiting in the pipeline, Scalpel needed its platform to grow with demand.
Scalpel's original architecture ran inference on a single instance, typically tied to fixed local hardware. Put simply, the models powering the app lived on physical servers that Scalpel owned and managed. That worked for a handful of models. It did not work for 300.
Unlike many computer vision systems that use one general-purpose model, Scalpel's approach requires a dedicated model for every unique tray configuration. With approximately 300 tray variants in scope, and no single machine capable of holding that many large models simultaneously, the single instance approach had reached a hard ceiling. The business had proven its concept and won its first major contract. It just needed an intelligent mechanism to serve all models easily and simply.
Three challenges shaped the engagement:
Scalability. Serving around 300 model variants from a single platform, with sub-10-second response times, was not achievable on-premise. The architecture needed to move to the cloud, where compute can expand and contract on demand.
Compliance. Scalpel operates in healthcare but had not yet defined and implemented a compliance framework. With HIPAA, ISO 27001, and SOC 2 all relevant to their client base, they needed to close that gap before any clinical trial could proceed with confidence.
Time pressure. A large healthcare provider trial was already booked; Scalpel AI needed the platform ready.
The Solution
Cloudscaler delivered the engagement in approximately five weeks, ahead of schedule, building two workstreams in parallel to keep the timeline tight.
A Healthcare-Aligned AWS Landing Zone
The foundation is a seven-account AWS Landing Zone, a purpose-built, governed cloud environment that separates development, test, and production workloads and embeds security and compliance controls from the ground up. Delivered entirely as Infrastructure-as-Code, meaning every configuration is version-controlled, auditable, and repeatable rather than set up manually, it established:
Centralised identity and access management, ensuring that every person and system can only access what they are permitted to, with multi-factor authentication enforced and all access permissions defined by role.
Segmented networking across dedicated development, test, and production accounts, keeping environments cleanly isolated so that a change in one cannot inadvertently affect another.
Encryption of data at rest and in transit, automated backups with cross-account disaster recovery, and data residency fixed to a single AWS region.
Continuous compliance monitoring aligned to HIPAA, ISO 27001, and SOC 2, with active checks running against the environment at all times and issues surfaced automatically rather than discovered in a manual audit.
A CI/CD pipeline, the automated pathway that takes a code change from a developer's machine through testing and into production, was configured with approval gates at each environment boundary, reducing the risk of untested changes reaching live systems.
SageMaker Multi-Model Endpoint
The surgical-tray models were migrated from on-premise hardware to AWS SageMaker, Amazon's managed machine learning service. Rather than provisioning a separate server for each of the 300 models (which would be prohibitively expensive and operationally unmanageable), Cloudscaler implemented a Multi-Model Endpoint: a single, shared serving layer that loads the relevant model on demand when a request arrives. An API Gateway acts as the front door, routing each incoming request from the mobile app to the right model and returning the result.
To validate performance, Cloudscaler ran 184 payload configuration tests against the on-premise baseline. A 4 MB payload change cut inference latency significantly with no measurable accuracy loss, giving Scalpel a faster, lighter solution than the one it had replaced.
Benefits
The immediate outcome is that Scalpel's platform works at scale. All ~300 surgical-tray models now run from a single, elastic endpoint that scales with demand. Latency is within the sub-10-second target. The system is ready for the trial and for the clients lined up behind it.
Beyond the inference capability, Scalpel now has a security and compliance posture appropriate for a healthcare technology business. Detective controls aligned to HIPAA, ISO 27001, and SOC 2 are configured and active. Audit logging, encrypted storage, and identity controls are in place across every account. The compliance framework that did not exist at the start of the engagement is now built into the platform.
Knowledge transfer was embedded throughout. Three structured sessions covered architecture, operations, and troubleshooting, with recordings made available for the Scalpel team. A four-week hypercare period followed handover to support the transition to self-sufficient operations.
The Value
The most significant thing Cloudscaler delivered was not the inference endpoint. It was the foundation underneath it.
Scalpel came into this engagement without a codified, governed, managed AWS environment. What they now have is a production-ready, healthcare-aligned Landing Zone that gives them a complete vertically sliced blueprint of what good looks like. Every control, account boundary, and compliance check is codified and repeatable. As Scalpel grows, whether that means onboarding new clients, expanding model coverage, or migrating additional workloads, they have an established, well-architected cloud operating model to build into rather than around.
Lessons Learned
Parallel workstreams protect the timeline. Building the Landing Zone alongside model discovery, rather than sequentially, kept a five-week delivery achievable and meant the compliance foundation was ready when the inference work needed it.
Data-driven decisions resolve technical debates faster. An image size payload test that identified the 4 MB optimisation settled a trade-off that could otherwise have been a blocker to the solution. Running the numbers early removed ambiguity and gave the team a clear, defensible path forward.
This is the kind of outcome we're proud to build alongside our customers, and there's more to come. Join us as we present alongside Scalpel AI at the AWS Public Sector AI Symposium London on 9 September, where we'll be sharing the story behind this project in person.