AI Blocks

What is an AI Block?

Units of AI that act like callable functions

Think of AI Blocks as ingredients and DPI Workflows as recipes. Each block does one bounded thing — translate, verify, extract, classify — and can be composed into any service workflow. The block is replaced; the workflow remains.

🎯

Minimalist

One clear purpose. Bounded inputs and outputs. No hidden side effects.

🧩

Composable

Interoperates with other Blocks and workflow engines through open APIs and common data formats.

♻️

Reusable

Applicable across agencies, sectors, and countries without rework. One block, many services.

🔍

Governable

Observable, testable, policy-aligned. Built-in audit trails, confidence scores, and privacy controls.

Classification

Two types of AI Blocks

🌐

Foundational AI Blocks

General-purpose capabilities reusable across all sectors. The multimodal and local language primitives that are prerequisites for an inclusive AI-ready nation.

translate() — local language translation & transliteration
speech_to_text() — voice input in any dialect
text_to_speech() — voice output accessibility
ocr_extract() — optical character recognition
summarise() — document summarisation
image_classify() — image and video recognition

⚙️

Sector-Specific AI Blocks

Tailored to specific workflows. Embed policy and operational logic into callable functions that serve public mandates.

eligibility_verify() — social protection eligibility
identity_verify() — digital ID verification
duplicate_detect() — civil registration deduplication
clinical_decision() — clinical decision support
personalized_tutor() — adaptive education
fraud_detect() — disbursement anomaly detection
area_classify() — GIS crisis area classification

Safeguards as Callable AI Blocks: Safeguards can themselves be AI Blocks — bias_check(), consent_verify(), anomaly_flag(). By treating safeguards as modular, replaceable components, governments can evolve their risk management without rebuilding workflows.

Block Specification

Every AI Block declares its contract

A block spec defines not just what the block does, but how it can be governed, replaced, and audited. The spec is machine-readable and published alongside the block.

ai_block:
  id: "eligibility_verify_v1"
  version: "1.0.0"
  description: "Verifies citizen eligibility for a named benefit program"
  type: sector_specific
  domain: social_protection

  interface:
    inputs:
      citizen_id:   {type: string, format: national_id, required: true}
      benefit_code: {type: string, enum: [CASH_TRANSFER, FOOD_VOUCHER], required: true}
      consent_token: {type: string, format: jwt, required: true}
    outputs:
      eligible:      {type: boolean}
      confidence:    {type: number, min: 0, max: 1}
      reason_code:   {type: string}
      audit_log_ref: {type: string}

  governance:
    confidence_threshold: 0.85
    on_low_confidence: human_escalate
    pii_fields: [citizen_id]
    pii_handling: tokenise_before_log
    audit_logging: true
    data_retention: "90d"

  provenance:
    model_type: "gradient_boosted_classifier"
    training_data: "social_registry_2023_anonymised"
    last_evaluated: "2025-12"
    hosting: sovereign_cloud         # sovereign_cloud | on_prem | external_api
    replaceability: drop_in_compatible  # can be swapped without workflow changes

Interoperability

Model Context Protocol (MCP)

As AI Blocks evolve into reusable capabilities, they need a common way to exchange context and interact safely. The Model Context Protocol (MCP) illustrates how this interoperability can be achieved — a lightweight protocol that allows AI systems to access external data, tools, and services while maintaining contextual integrity and auditability.

Within the DPI-AI Framework, MCP-inspired protocols enable:

AI Blocks to communicate and collaborate through shared context layers
Similar to how REST APIs enable interaction across other DPI building blocks
Composable, transparent, and portable AI capabilities across ecosystems
Governments can ensure AI capabilities remain interoperable without vendor lock-in

# MCP-style AI Block invocation
POST /mcp/v1/invoke
{
  "block": "eligibility_verify_v1",
  "context": {
    "session_id": "wf-2026-001",
    "consent_token": "eyJ...",
    "audit_trail": true
  },
  "inputs": {
    "citizen_id": "KE-1234-5678",
    "benefit_code": "CASH_TRANSFER"
  }
}

→ See which DPGs are AI Block / MCP ready