Artificial Intelligence Subject Matter Expert (CBP

About Agile Defense

At Agile Defense we know that action defines the outcome and new challenges require new solutions. That’s why we always look to the future and embrace change with an unmovable spirit and the courage to build for what comes next.

Our vision is to bring adaptive innovation to support our nation's most important missions through the seamless integration of advanced technologies, elite minds, and unparalleled agility-leveraging a foundation of speed, flexibility, and ingenuity to strengthen and protect our nation’s vital interests.

Title:

Artificial Intelligence Subject Matter Expert (AI SME) (CBP)

Clearance:
Active CBP Background Investigation (CBP BI) and EOD strongly preferred. We can begin processing for candidates who do not hold one.

Citizenship: U.S. Citizenship required

Location:

Ashburn, VA

Salary Range: [Pending]

Signing Bonus: $10,000 for candidates with an active CBP BI. Payable after 90 days; standard terms apply.

Travel:
Rare, as needed

U.S. Customs and Border Protection runs continuous operations across more than 300 land, air, and sea ports of entry, plus Border Patrol stations and the Air and Marine Operations Center. Applying artificial intelligence in that environment is not the same problem as applying it somewhere the cost of a wrong answer is a bad recommendation. A model that runs against operational data here has to be accurate, explainable, and safe under federal accreditation, or it does not ship, no matter how well it performs in a lab.

You are the technical authority on what that requires. You will design, test, and align AI and machine learning models so they perform reliably in a professional operational setting, working alongside program leadership, engineering teams, and the security staff who have to sign off on what you build before it runs against real data.

Two things are worth knowing before you apply. This role exists to bring rigor, not to chase what is newest. And a meaningful part of the job is being the person in the room who can tell the difference between a model that looks impressive in a demo and one that is actually safe to run.

Objective 1:
Design models that are accurate enough to trust with real decisions

• Model performance gets validated against real operational conditions, not only against a clean benchmark dataset.
• Where a model is wrong, you can say how it is wrong and how often, rather than reporting a single accuracy number and stopping there.
• You can explain a model's behavior to someone who did not build it, in terms they can evaluate.

Objective 2:
Keep models safe and aligned once they are running against real data

• Model behavior gets monitored after deployment, not just validated once before it.
• Drift and degraded performance get caught before they change an operational outcome.
• Failure modes are documented and understood before a model goes live, not discovered afterward.

Objective 3:
Get AI capability through federal review without the review becoming a guessing game for everyone else

• Documentation and evidence for a security or accreditation review come out of your normal development process.
• You can tell an engineering team early what an AI capability will need to clear review, before they have built around an assumption that will not survive it.
• Reviewers get a straight answer about what a model does and does not do.

Objective 4:
Build technical judgment into the program that outlasts any one model

• Other engineers and program staff come to you before committing to an AI approach, not after it has already been built.
• Where AI is not the right tool for a problem, you can say so and be heard.
• Standards for evaluating and deploying models exist and get used on the next project, not reinvented each time.

Preferred Experience

• You have taken a machine learning model from prototype to a production system that real people depended on.
• You have worked in a regulated or federal environment where a model's decisions had to be explainable and defensible, not only accurate.
• You have caught a model failure before it reached a decision that mattered, and can describe how.
• You are comfortable across the model…