Dimension A: Intelligence layer of the Trust Graph
             Trajectory analysis, contextual interpretation, cross-track synthesis

Dimension B: Witnessing model scale engine
             Submission triage, consistency checking, regional pattern learning

Dimension C: Resident advocate intelligence layer
             Supports decisions across all three Resident tracks

Dimension D: Professional operator intelligence layer
             Supports agents, principals, organisations, and formal operators

Dimension E: Network-level fraud detection
             Velocity anomalies, geographic clustering, cross-party patterns

Dimension F: Federated computation engine
             Bank models run on Leja data — raw data never leaves

Dimension G: Service participation trust matcher
             Trust compatibility for service, space, and stay participation

apps/backend/src/modules/intelligence/
├── intelligence.module.ts          NestJS module — exported to all other modules
├── intelligence.service.ts         Core: model selection, rate limiting, cost tracking
├── extraction/
│   ├── extraction.processor.ts     BullMQ worker
│   ├── extraction.prompts.ts       Bank-specific prompts (GTB, Access, UBA, etc.)
│   └── extraction.schemas.ts       Zod schemas for generateObject output
├── document-parser/
│   ├── parser.processor.ts         BullMQ worker
│   ├── parser.prompts.ts           Agreement parsing prompts
│   └── parser.schemas.ts           Zod schemas
├── narrative/
│   ├── narrative.service.ts        Sync service
│   └── narrative.prompts.ts        Nigerian English register prompts
├── anomaly/
│   ├── anomaly.processor.ts        BullMQ scheduled worker
│   └── anomaly.schemas.ts          Zod schemas for anomaly output
├── prediction/
│   ├── prediction.processor.ts     BullMQ scheduled worker
│   └── prediction.schemas.ts       Zod schemas
├── fraud/
│   ├── fraud.processor.ts          BullMQ worker (post-payment)
│   └── fraud.schemas.ts            Zod schemas
├── query/
│   ├── query.service.ts            Sync service (NL → filter)
│   └── query.prompts.ts
└── shared/
    ├── model-selector.ts           Routes to correct model by task
    ├── cost-tracker.ts             Per-org AI cost tracking
    ├── retry-strategy.ts           Exponential backoff, fallback switching
    └── prompt-utils.ts             Template helpers, context builders

import { generateObject } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';
import { z } from 'zod';

const PaymentSignalSchema = z.object({
  amount: z.number().positive(),
  paymentDate: z.string().date(),
  reference: z.string().optional(),
  payerName: z.string().optional(),
  institution: z.string().optional(),
  rail: z.enum(['BANK_TRANSFER', 'CARD', 'USSD', 'VIRTUAL_ACCOUNT', 'CASH', 'OTHER']),
  confidence: z.number().min(0).max(1),
  reviewNotes: z.string().describe('Reasoning for confidence level'),
});

const { object } = await generateObject({
  model: anthropic('claude-sonnet-4-5-20251001'),
  schema: PaymentSignalSchema,
  messages: [{
    role: 'user',
    content: [
      { type: 'image', image: normalizedEvidenceUrl },
      { type: 'text', text: PAYMENT_REVIEW_PROMPT },
    ],
  }],
});
// object is fully typed and validated — no parsing needed

import sharp from 'sharp';

async function preprocessPaymentEvidence(buffer: Buffer): Promise<Buffer> {
  return sharp(buffer)
    .greyscale()
    .normalize()                    // Auto-contrast
    .sharpen({ sigma: 1.5 })       // Sharpen text edges
    .resize(1200, null, { withoutEnlargement: true })
    .png()
    .toBuffer();
}