A county government agency hosting an annual essay contest for K–12 students anticipated a high volume of submissions requiring pre-screening before human graders could begin scoring. The grading pool is a finite resource. Routing ineligible submissions — essays that fail to meet basic prompt requirements or that show signs of AI generation — to human graders wastes time that should be spent evaluating essay quality. A manual pre-screen at scale is not operationally realistic.
The requirement was a fast, consistent, and scalable pre-filter that could be applied uniformly to every submission regardless of volume, while keeping human graders in control of all final quality judgments.
The design was organized around a single governing principle: each architectural layer owns what it does best. The submission form owns structured data validation. The serverless compute layer owns server-side validation logic independent of the client. The AI layer owns content semantics — the questions that require reading and understanding the essay, not just parsing it.
Three decisions shaped the implementation in ways that matter specifically for government and K–12 contexts. First, no student personally identifiable information (PII) is forwarded to the AI model — an explicit architectural decision made with the Family Educational Rights and Privacy Act (FERPA) in mind. Second, AI suspicion of AI-generated content triggers a mandatory human review flag, not an automatic disqualification — human reviewers make all final eligibility determinations. Third, grade-band calibration was built into the AI prompt, evaluating each submission against what is developmentally appropriate for that grade band rather than a single adult-writing standard.
The solution is a serverless, event-driven pipeline deployed entirely within Microsoft's government-authorized cloud infrastructure. Azure resources — the Function App, Azure OpenAI Service, and Blob Storage — run in Azure Government (US Government Virginia region, FedRAMP High authorized). Dataverse runs in the Power Platform Government Community Cloud (GCC) environment. The two environments are connected via OAuth 2.0 client credentials authentication, with the Azure Function authenticating to Dataverse as a service principal scoped to write access on the submission table only.
Students submit essays through a purpose-built form hosted as a static website on Azure Blob Storage. Each submission triggers an Azure Function built on .NET 8 using the isolated worker model. The Function performs independent server-side word count validation, resolves the student's grade level to one of four grade bands, and constructs a prompt for Azure OpenAI Service. The model returns a strict JSON response containing eligibility status, individual criterion results, AI-generation suspicion flag, and human review requirement flag. The Function then writes the complete submission record and screening result to Dataverse via the Web API. The entire pipeline — from student submission through AI screening to grader-ready record — executes within seconds.
Government agencies regularly host public-facing programs — contests, grant applications, scholarship submissions, program enrollment — that generate high-volume intake requiring pre-qualification before human review. This engagement demonstrates that AI-powered intake screening is deployable within government cloud boundaries, compliant with student data privacy requirements, and architecturally defensible — with human review preserved for every edge case the AI cannot resolve with confidence. The pipeline architecture is repeatable across grant pre-screening, scholarship eligibility filtering, program application intake, public comment classification, and any scenario where AI-assisted pre-qualification can protect finite reviewer capacity.
Schedule a 30-minute conversation about your operational challenges.
ore.brigham@brigville.com