Data lifecycle
Identify authoritative records, history requirements, update frequency, retention, and the consequences of stale information. Preserve source event time and ingestion time where ordering matters.
INDUSTRY ARCHITECTURE / Public sector
Public sector architecture
A connected reference design spanning domain information, data engineering, AI services, agent execution, and governance.
DOMAIN DESIGN DECISION
The business context
shapes the architecture.
Public service guidance and internal case information have different audiences, and document completeness does not determine eligibility.
Preserve procedure versions, separate public and restricted sources, and represent missing information explicitly.
Start with citizen information rather than attempting to automate the whole industry at once. Define a bounded request, the evidence required to answer it, and the person accountable for accepting the result.
DATA → AI → AGENTIC AIDOMAIN CONTEXT: Service / Application / Required document / Procedure version
Data architecture
- 01
Domain sources
Published procedures, application documents, case correspondence. Identify the system of record, owner, refresh rate, and approved purpose for each source.
- 02
Ingestion & contracts
Batch, change-data capture, events, and document intake as appropriate. Validate schema, identity keys, event time, and rejected records.
- 03
Storage & transformation
Separate raw evidence, validated records, and curated domain products. Preserve version history and reconciliation results.
- 04
Semantic serving
Service · Application · Required document · Procedure version. Publish agreed definitions, access rules, and reusable interfaces for analytics and AI.
AI architecture
- 01
Approved information
Use curated domain products and approved documents. Preserve metadata, effective dates, and access scope.
- 02
Retrieval & features
Build permission-aware retrieval indexes or model features for the specific task. Test relevance and freshness independently.
- 03
Model & evaluation
Evaluate candidate methods on representative domain examples. Check groundedness, completeness, uncertainty, cost, and failure behavior.
- 04
Experience & feedback
Serve sourced drafts or scored signals through a review experience. Capture feedback without silently retraining or widening data access.
Agentic AI architecture
- 01
Task & identity
Record the initiating user, allowed purpose, task boundary, and completion criteria.
- 02
Reasoning & state
Citizen information agent and Case preparation agent. Keep task state, retrieved evidence, and unresolved questions traceable.
- 03
Scoped enterprise tools
Expose typed, permission-checked search, lookup, and draft tools. Validate arguments and handle retry or timeout safely.
- 04
Approval & audit
Eligibility, enforcement, and other consequential decisions remain with accountable officials. Record tool calls, reviewer decisions, and the accepted result under an agreed retention policy.
Shared governance & operationsOwnership · classification · lineage · access · quality · retention · evaluation · observability · incident response · cost management
Implementation decisions
to resolve before build.
Serving & integration
Choose batch, event, API, or document access based on workload needs. Separate read-only retrieval from tools that alter enterprise records. Define timeouts, retries, and recovery ownership.
Operational evidence
Measure checklist accuracy, source-version correctness, access-boundary testing, and review packet completeness. Keep versioned evaluation cases, end-to-end traces, and reviewer feedback so changes can be assessed against an agreed baseline.
CASE
A public-service application preparation blueprint
See the proposed architecture applied to a focused workflow.
Read more ↗PRODUCT BLUEPRINT
Public service knowledge workspace
Organize approved public information and casework preparation.
Read more ↗INDUSTRY HUB
Public sector
Explore domain workflows, products, agents, and related knowledge.
Read more ↗A FOCUSED CONVERSATION
Connect the architecture
to your priorities.
Discuss the domain, sources, controls, and outcomes that would define a useful first engagement.
Discuss this topic ↗