Learning Objectives

After completing this lecture, participants will be able to design, plan, and socialize a federal agency-wide AI platform that consolidates data, model, inference, monitoring, and governance capabilities while complying with OMB Memorandum M-24-10, the NIST AI Risk Management Framework 1.0, FISMA, FedRAMP, Section 508, Executive Order 14028 on software supply chain, and Executive Order 14110.

Specifically, you will be able to:

1. Describe the reference architecture of an agency-wide AI platform: data ingestion and curation, feature and vector stores, training compute, model registry, inference serving (including retrieval-augmented generation and agentic workflows), evaluation and red-teaming tooling, monitoring and observability, and AI governance workflows.
2. Apply the build-versus-buy-versus-share decision framework, including shared services from the General Services Administration (GSA), the Centers of Excellence at GSA's Technology Transformation Services, the National AI Research Resource pilot, DoD's CDAO platforms (Advana), HHS's AI platform at NIH, and the VA's AI Tech Sprint program, so the agency does not rebuild what a federal shared service already offers.
3. Design the authorization boundary of the platform so it obtains and maintains a FedRAMP-appropriate Authority to Operate under FISMA with AI-specific overlays consistent with the NIST AI RMF and MITRE ATLAS.
4. Organize the platform's governance model so that every program using the platform inherits an AI Governance Board-approved baseline while retaining program-level accountability under M-24-10.
5. Apply the Technology Modernization Fund (TMF) investment model where applicable, align to FITARA and Clinger-Cohen investment review, and integrate with the agency's Enterprise Architecture.
6. Integrate with agency identity and access management using Personal Identity Verification (PIV) consistent with HSPD-12 and FIPS 201, and with Zero Trust Architecture per NIST SP 800-207 and OMB M-22-09.
7. Incorporate Section 508 accessibility into all citizen-facing surfaces, with testable acceptance criteria.
8. Communicate the platform plan to the Chief Information Officer, Chief AI Officer, Chief Data Officer, Chief Information Security Officer, Senior Agency Official for Privacy, Chief Acquisition Officer, and the agency head, in terms each can approve.

Key Topics Covered

Topics covered in this lecture include:

Reference architecture. Data ingestion; data-quality and privacy preprocessing; feature stores; vector stores for retrieval-augmented generation; training compute (in-agency or via the National AI Research Resource); model registry; inference serving (real-time, batch, agentic); evaluation and red-teaming infrastructure; monitoring and observability; and governance workflows (intake, review, approval, inventory, lifecycle tracking).

Build vs. buy vs. share decisions. Where the General Services Administration, Federal Chief Information Officer Council, and 18F provide shared services. Where DoD's Advana provides platform baseline. Where HHS, DHS, or VA provide sector-specific baselines. Where the National AI Research Resource pilot (NAIRR) provides training compute to federal researchers. Where commercial FedRAMP offerings are the right default.

Authorization boundary. Single FedRAMP ATO for platform-level controls, with program-level inheritance and program-specific controls on top. How to use the FedRAMP PMO, JAB, or agency-ATO path.

Multi-tenant governance. How one platform supports many program offices without cross-contamination, with tenant-level controls, data isolation, per-tenant logging, and per-tenant exit rights.

Identity and Zero Trust. PIV-based authentication for human users; FIPS 201 for federal PKI; ICAM integration; attribute-based access control; machine-to-machine authentication via mTLS and OIDC; Zero Trust Architecture per NIST SP 800-207 and OMB M-22-09.

M-24-10 platform-level commitments. How a platform codifies minimum practices for rights-impacting and safety-impacting use cases once, so program offices inherit them rather than reinvent them.

NIST AI RMF operationalization. How the platform's tooling embodies Govern (governance workflows), Map (use-case inventory intake), Measure (evaluation harnesses and monitoring), and Manage (incident response and risk treatment).

Data management. Federal records (Federal Records Act 44 U.S.C. 3301), Privacy Act 5 U.S.C. 552a, sector-specific data (FTI, PHI, CUI, CJI), data-sharing agreements, and data-use rights in contracts.

Observability and Continuous Diagnostics and Mitigation (CDM). How platform telemetry feeds the agency SOC, CISA CDM, and the AI Governance Board.

Accessibility. Section 508 compliance for every citizen-facing surface on the platform, with automated testing.

Investment and planning. FITARA, Clinger-Cohen, TMF, appropriations cycles, and Enterprise Architecture alignment.

Worked example. A mid-size agency's platform roadmap from Year 0 (baseline assessment) to Year 3 (steady-state operation).

Why This Matters for Government

Federal agencies that adopt AI one program at a time end up with sprawl. Program offices procure their own inference services, build their own monitoring, sign their own contracts, obtain their own ATOs, and produce their own evaluation evidence. The agency has no common view of its AI use-case inventory, cannot answer OMB M-24-10 compliance questions coherently, pays multiple vendors for overlapping services, and cannot enforce minimum practices uniformly. This is the pattern the Government Accountability Office has observed repeatedly, and it is the pattern the Office of Management and Budget M-24-10 sought to end by requiring an enterprise approach.

An agency-wide AI platform is the architectural response. By consolidating data, compute, model, inference, evaluation, monitoring, and governance tooling onto a single shared platform, the agency can: enforce a single ATO baseline; enforce M-24-10 minimum practices once; produce one coherent use-case inventory; negotiate enterprise-scale contracts; and give program offices a fast path from idea to production without each program relearning federal compliance.

The federal government has direct experience with this pattern. The Department of Defense's Advana platform, operated by the Chief Digital and Artificial Intelligence Office (CDAO), provides a data and analytics baseline across DoD components. The Department of Health and Human Services operates AI platforms at the National Institutes of Health, the Centers for Medicare and Medicaid Services, and the Food and Drug Administration. The General Services Administration's Technology Transformation Services offers shared services (18F, Centers of Excellence) that many agencies leverage. The National AI Research Resource pilot provides compute and data access to federal researchers without each agency buying GPUs. The Department of Veterans Affairs operates AI infrastructure across Veterans Health Administration that other VA components use. The Department of the Treasury operates analytics platforms across bureaus. The pattern of enterprise platforms, rather than one-off program procurements, is already well established across the federal enterprise and has been validated by GAO cost and schedule analyses.

The downside of doing this badly is also well documented. Platforms that are built without tenant governance become brittle; program offices exit to shadow IT. Platforms that are built without performance monitoring become unused; they miss AI-specific failure modes. Platforms that are built without the Chief AI Officer's sign-off sprawl into inconsistent minimum-practice implementations. Platforms that are built without Section 508 from day one create accessibility debt that blocks citizen-facing deployments later.

This lecture, therefore, gives practitioners a reference architecture they can pressure-test, a governance model that maps to M-24-10, an authorization strategy that uses FedRAMP rather than invents a new one, a shared-services posture that avoids rebuilding what the federal enterprise already offers, and an investment model that aligns to FITARA and the Technology Modernization Fund. The outcome is a platform that the Chief Information Officer, Chief AI Officer, Chief Information Security Officer, Chief Data Officer, Senior Agency Official for Privacy, Chief Acquisition Officer, and agency head can each approve because it solves their concerns explicitly.

WHY THIS MATTERS FOR GOVERNMENT

An agency-wide AI platform is not merely a technology project. It is the agency's AI factory. The design choices made at the platform level become the defaults for every future program AI deployment. Getting these right is a ten-year decision.

REFERENCE ARCHITECTURE AT A GLANCE

Data layer. Ingestion from agency systems of record, quality checks, privacy preprocessing, data catalog, access controls, data-use agreements surfaced as policy, retention under the Federal Records Act. Includes a feature store for structured ML and a vector store for retrieval-augmented generation.

Training compute. GPU or accelerator capacity either in the agency (usually a FedRAMP High tenancy) or via the National AI Research Resource pilot or the DOE's leadership computing facilities (Oak Ridge, Argonne, Lawrence Berkeley) as appropriate. A model registry tracks every trained artifact with lineage.

Inference serving. Real-time APIs for interactive use, batch APIs for offline, and agentic orchestration with tool calling. Scoped tool permissions by tenant. Retrieval-augmented generation for citizen-facing knowledge tasks. Rate limits. Telemetry.

Evaluation and red-teaming. Harnesses for accuracy, fairness, robustness, adversarial, prompt-injection, indirect-prompt-injection, model-theft, and agentic-tool-abuse probes. Findings feed back into pre-deployment testing and ongoing monitoring.

Monitoring. Dashboards for accuracy, fairness, drift, usage, tool calls, and incidents. Alerts. Incident response. Integration with agency SOC and with CISA's CDM.

Governance. Intake form for new use cases. Risk-tier classification. Minimum-practice checklist inheritance. AI Governance Board review. Approval workflow. Use-case inventory publication.

AUTHORIZATION BOUNDARY

The platform obtains a FedRAMP-appropriate Authority to Operate. Program tenants inherit platform-level controls and add program-specific controls. This is the pattern already used for federal shared services. The platform documents control inheritance so program ATO work is lighter, not lighter plus missing.

IDENTITY AND ZERO TRUST

PIV-based authentication (HSPD-12, FIPS 201) for human users; ICAM integration with agency identity; attribute-based access control; machine-to-machine via mTLS and OIDC; micro-segmentation per NIST SP 800-207 Zero Trust Architecture and OMB M-22-09.

BUILD VS. BUY VS. SHARE

Before investing, survey federal shared services. GSA Centers of Excellence; DoD Advana (for DoD); HHS platform offerings; National AI Research Resource; leadership computing facilities; commercial FedRAMP offerings. Share where a federal option meets the need. Buy where a commercial FedRAMP option meets the need. Build only what is agency-specific.

INVESTMENT AND PLANNING

Align with FITARA investment review, Clinger-Cohen requirements, Enterprise Architecture, and the Technology Modernization Fund where applicable. The platform is a major IT investment; it is reviewed by the CIO, by OMB/OIRA through the Capital Planning and Investment Control process, and by the Governance Board.

ACCESSIBILITY

Section 508 compliance is a platform-level commitment. Every citizen-facing surface is tested with assistive technology. Accessibility debt is tracked. Remediation is funded.

The rest of this lecture walks the reference architecture section by section with implementation guidance.

Related Lectures

L3 3.5.1 -- Cross-Agency AI Collaboration

L3 3.5.3 -- Emerging Technologies and Their Implications

L3 3.6.1 -- Supply Chain Risk Management for AI

L3 3.6.2 -- FedRAMP for AI Services

L3 3.7.1 -- Enterprise Architecture Alignment for AI

L3 3.7.2 -- Investing in AI Using the Technology Modernization Fund

L2 2.2.2 -- Cybersecurity for AI Systems

L2 2.2.3 -- Privacy Engineering for AI

L4 4.1.x -- Advanced Platform Operations

L4 4.2.x -- Multi-Tenant Governance for Federal Platforms