The System Design Round
Explore designing a document processing system for a public sector agency, focusing on handling diverse document types, ensuring data is processed in-country, and maintaining audit trails. Learn to sketch workflows, manage document quality, and apply rigorous completeness and review checks. Understand how to defend architectural decisions with evidence and address regulatory requirements while ensuring efficiency and accountability.
We'll cover the following...
- Sketch the pilot
- Receive and split A17
- Turn extraction into intervention
- Review every critical field
- Reconstruct the incomplete flag
- Catch confident income errors
- Measure difficult inputs
- Locate every A17 operation
- Evaluate Document AI first
- Block an impermissible extractor
- Price 4.8 million pages
- Size deadlines and changes
- Defend the revised architecture
Give the agreed pilot twenty minutes of design work before choosing an extractor. The difficult result is a prepared case whose values, completeness findings, and corrections remain explainable through changing evidence. Your architecture must also show where every citizen-data operation executes, because a useful extraction cannot be used if its processing path crosses the agency’s boundary.
Sketch the pilot
Try it first : (20 minutes) Sketch the one-application-type, one-regional-office pilot. Cover document flow from arrival to caseworker; how workers know which fields to trust; sovereign deployment; and the evidence explaining an incomplete flag to an auditor. Record unresolved service-location questions without inventing approvals or acceptance thresholds.
Create your initial architecture and open-question list before reading the worked design.
Keep the sketch available for revision. A box labeled “human review” needs an evidence interface, and a box labeled “in-country cloud” needs processing-location evidence for its dependencies.
The fictional Housing Assistance Agency handles about 400,000 applications annually, 60% on paper, with seven form versions from the last 15 years and limited legacy-system integration. Caseworkers retain eligibility decisions, and applicants retain appeal rights.
The interviewer adds packet A17: A form, ID, payslips, tenancy agreement, and utility-evidence photo. The photo is unreadable. Declared income is 1,800; the payslip visibly shows 1,800, but extraction suggests 7,800 with confidence 0.99. A clearer utility attachment arrives later.
Receive and split A17
Discovery examined 500 applications: 23 document types, 18% handwritten forms, about 10% poor scans or phone photos, and about 30% incomplete applications. Most elapsed time was waiting for missing documents. These sample findings motivate early completeness assistance; they do not establish exact national subgroup counts.
The worked flow begins with two stages:
Intake quality checks: Accept scans, uploads, and phone photos. Preserve the originals, receipt time, packet identity, and received-page inventory. Check blur, orientation, and expected pages where expectations are known. Request an immediate re-upload of A17’s unreadable photo through the approved workflow.
Classification and splitting: Separate the mixed packet into document units labeled by type. Preserve the mapping from each unit back to its original file, revision, and page. Route uncertain classifications for inspection instead of silently assigning a convenient type.
Distinguish three observations:
Known missing page: An established page expectation is unmet.
Unreadable page: Material arrived, but its contents could not be inspected adequately.
Unknown composition: The workflow cannot yet establish what documents the packet contains.
For A17, retain the utility photo and its failed-readability outcome. The justified claim is “utility evidence inspection incomplete.” Neither that observation nor a missing-page finding establishes ineligibility.
Turn extraction into intervention
Continue in processing order:
Structured extraction: Populate an agreed schema, retaining each suggestion’s source anchor, source revision, extraction version, and field confidence. A source anchor identifies the precise page location supporting the suggestion.
Cross-document validation: ...