Machine learning integration for existing applications
Connect prediction and classification capabilities to the applications and decisions they need to support. We work with your model specialists or provider on the surrounding software, defining data flows, failure handling and user review so the integration can be operated and maintained.
What you can expect
- Defined data flows and model responsibilities
- A tested application and API integration
- Fallback, monitoring and handover arrangements
Senior specialists in engineering, design and delivery.
How we assure qualityWhat we can help with
- Data-source and model API connections
- Background processing and result handling
- Application interfaces for model outputs
- Integration testing and failure recovery
- Monitoring and operational handover
Turn a model output into a usable service
A prediction has to reach the right person or process at the right time. The surrounding software must supply suitable inputs, handle missing results and make it clear when a person needs to intervene.
Our role is application engineering and integration. We can work with your internal specialists or a model provider to connect their capability to an existing product. Any requirement for model research, training or specialist validation needs to be established separately when we define the scope.
Define the responsibilities around the model
An integration starts with questions that affect the whole service:
- Which application owns the source data, and what may be shared with the model?
- Who is responsible for assessing the model’s suitability and the quality of its outputs?
- How will the application handle unavailable, late or incomplete results?
- Which decisions require review, and how can a user correct or challenge an output?
- Who will monitor the service and investigate changes in behaviour?
These questions help separate application reliability from model quality. Both matter, and they may have different owners.
Integration experience: Microsoft SwiftKey
For Microsoft SwiftKey, we built a Rails backend that connected users’ permitted data from other services to SwiftKey’s personalised prediction engine. We used JRuby to work with the client’s Java libraries and existing deployment environment, and agreed the API with the mobile development team.
This was backend and API integration around SwiftKey’s prediction technology. The case study does not claim that we trained its model. It shows the work needed to connect a specialist capability to an existing product and team.
Scope the engineering work
Depending on the application, the work may include data-source connections, API design, background processing, user interfaces, integration tests and operational monitoring. We agree ownership and handover alongside the implementation so that the service can be supported after launch.
See our technical assurance for the wider delivery practices. For language-model applications such as document search or assisted drafting, use generative AI development. If you are still deciding whether AI is appropriate, AI consulting and automation is the starting point.
Experience in practice
Relevant work
Microsoft Swiftkey
Building an API that linked Microsoft SwiftKey with individual social media and email accounts.
In three months, we built an API that links Swiftkey with individual social media and email accounts. Each user now experiences a completely personalised service.
Technology
Plan your next step
Bring the application, proposed model or prediction task, and the people who will use the result. We can assess the integration boundaries and what still needs to be established.
