Other Azure components included Application Insights for logging and reporting, and the AI Computer Vision instance itself. In the main, our teams developed these components using C# .Net Core in Linux, and then secured all components using IP address/active directory permissions via the Azure portal.
esynergy teams utilised Microsoft Azure AI services across numerous applications when building this tooling. For example, we used Azure Computer Vision to develop a prototype handwriting review portal, which can intelligently analyse handwritten text in image files then provide structured results for users to review.
Our approach: Agile, visible, collaborative
esynergy used agile working methods and automation to maximise efficiency and reduce costs throughout this project.
For instance, we developed Terraform scripts to build and configure the required Azure components. Website deployment and associated function code was also automated.
Visibility was key. We used Azure DevOps to manage stories, and ran regular stand-ups, planning sessions, progress reviews and demos to keep teams and stakeholders informed and involved. We created a clear digital paper trail of progress, setting out how decisions were made and changes were managed.
Our team kept the client’s legal staff involved throughout the journey, which one of the client’s technical specialists called “massively collaborative”.
For example, we created an initial test harness that allowed users to compare AI services, including stable and beta versions of Azure Computer Vision. Users could upload a document to multiple different cloud offerings, then use the results to decide which one to incorporate in the prototype portal.
At every stage, we consulted and tested with the client’s legal professionals to find out what was working well, and what needed improving. Elements that proved particularly popular included the ability to load a case very quickly, switch quickly between multiple documents in tabs, and carry out redaction on the fly.
The tool’s instinct for grouping cases together was harder to get right, because categorisation is subjective. AI and machine learning will improve the tool’s grouping decisions in time. Meanwhile, automated grouping still represents a big step forward in helpfulness and efficiency for users.
Next steps: From alpha to pilot
The project is now a fully working alpha-stage prototype running on dummy data. In fact, the system is running so smoothly that we’ve actually had to explain to users that the data isn’t real.
Lawyers, who will make up the bulk of the portal’s users, now account for about 90% of alpha testers. We have consulted them at every stage of the process, and asked them what they need to suit their ways of working.
We’re now in the process of switching to anonymised real legal data, with pilot testing due to begin in autumn 2022.