ECS experts speak to the importance of having a holistic understanding of AI implementation, from development to deployment
One of the worst things an organization looking to implement an AI solution can do is rush. The saying “what you put in is what you get out” is precisely applicable for AI, and if you rely on junk data or lack the means to validate model performance, then the information you base your operations on could be of dubious value. Whether you work in defense, healthcare, or at a federal law enforcement agency, poor implementation of AI can have disastrous real-world impacts.
We sat down with Chris Snow, principal software engineer of Mission Solutions at ECS, and Charles Stevens, program manager for Mission Solutions, to learn about what they call “deployable AI” — AI that has been proven to perform at scale before deployment to the field. Our experts also touch on the many pitfalls facing organizations looking to implement AI solutions, not all of which are obvious, as well as how ECS stands out as an integrator of AI solutions.
ECS is the leader in AI research, development, and deployment for the Department of Defense (DoD). Our experts have deployed more operational AI than any other DoD provider.
CHRIS SNOW
Principal Software Engineer,
Mission Solutions
CHARLES STEVENS
Program Manager,
Mission Solutions
Q: What are some of the most common problems organizations run into when it comes to effective use of AI?
Chris Snow: What people often don’t realize is just how multifaceted the challenges around data and AI really are. Just speaking from the data perspective alone, what we always try to communicate to our customers is that they want to achieve data-driven decision making. With that comes a wide array of challenges: having the knowledge and expertise to assess data quality, understanding your data pipeline and where this information is coming from and how it’s being processed at every stage, being able to conduct a data discovery process so you know what your organization has in its possession, determining who in the organization should have access to what data. These are all challenges that have a direct impact on the efficacy of your AI solutions, because the model only works if the data policy is sound.
Charles Stevens: Chris touched on this briefly, but another critical dimension to this is having a workforce with the skills to both use AI effectively and mitigate the risks, such as sensitive data loss. Do you have the expertise to judge what tools best serve your mission objectives? Has your organization clearly defined what constitutes acceptable AI use? Or, from an external threat perspective, are you following cybersecurity best practices to protect your tools from hacking and misuse? These are the questions every organization using an AI solution will have to answer fairly early on.
Q: How do we arrive at “deployable AI?” What is it about AI development, integration, and deployment that makes it a challenge to get an optimized model solving real problems in the field?
Chris Snow: One of the biggest challenges with AI development is validating model performance at scale, and there are several reasons for that. AI systems have huge computer processing and storage requirements as they scale up. They grow in technical complexity and the tasks of data cleansing and preparing test data can be time-and-labor intensive.
ECS has developed an automated pipeline that allows model vendors to rapidly test their models against a wide variety of data, without human intervention. We reduce the test-train-improve cycle time from days or weeks to a matter of minutes. This rapid turnaround of model performance results allows more models to be tested, resulting in more performant models reaching the field. That’s what we mean by deployable: the models are proven to perform at scale with a process that is automated, rapid, and repeatable.
Charles Stevens: It’s important to remember that effective AI deployment is an integration effort rather than a bolt-on construct. When it comes to deploying AI, the most significant challenges generally relate to multi-organization deployments wherein different user groups are employing different pre-existing technologies (both hardware and software) which may not be compatible.
Another major challenge is user adoption, which is most often related to a perception that “somebody, somewhere built something that nobody here needs or wants or asked for, because somebody, somewhere else thought it was a good idea.” You have to secure user buy-in.
Q: What sets ECS apart from other service providers when it comes to AI integration?
Chris Snow: A. ECS has a holistic understanding of AI model development, from training and validation to fielding models that solve real-world problems. We’re better at integrating AI because, again, of our automated development process. We offer vendors a complete environment where they can bring their own tools to train models against government data, as well as cross-domain solutions that allow them to transfer their models to classified environments for further enrichment.
ECS also has personnel in the field, working with customers to deploy and evaluate models against real-world problems. The feedback provided by field support is fed back to us to further improve models, enabling an iterative approach.
Charles Stevens: Within its Field Service Operations organization, ECS fields an experienced staff of intelligence professionals who work closely with user groups to integrate AI around the globe, including hazardous duty locations. Through a highly selective recruitment process, we onboard individuals chosen not only for their technical subject matter expertise, but also for their personal suitability for the particular site and organization they will support. Every member of the team approaches solutions in the field with innovation and imagination, rooted in a solid foundation of practical experience and a deep understanding of the users’ processes, procedures, and mission objectives. We don’t send coders to the field to just sit around and wait for someone to complain about a feature.
Our collaborative mindset also ensures user feedback enters the development and engineering loop nearly instantaneously, meaning users often see their operational needs being addressed faster than they’ve ever been accustomed to.
