How Pairing Jev with AWS Strands SDK Cut Our WhatsApp Agent Latency by 73% and Triage Costs by 99%

                                Inbound Parent WhatsApp Message
                                                │
                       ┌────────────────────────┴────────────────────────┐
                       ▼                                                 ▼
             [Approach A: Regexes]                             [Approach B: Generative LLM]
             - Latency: <10ms                                  - Latency: 1,500ms – 2,500ms
             - Cost: $0.00                                     - Cost: High token overhead
             - Fragility: EXTREME                              - Formatting: Regex JSON parsing
             - Breaks on typos, slang & dialects               - No calibrated confidence

                            Parent WhatsApp Message
                                       │
                                       ▼
                        [PII Redaction & Normalization]
                                       │
                                       ▼
               ┌───────────────────────────────────────────────┐
               │    SYSTEM 1: TypeSafe Jev (Sub-400ms)         │
               │  - Intent Triage: choice(11 categories)       │
               │  - Language Detection: choice('en' | 'fr')    │
               │  - Curriculum Completeness: noul(missingYear) │
               │  - Calibrated Statistical Confidence (0 - 1)  │
               └───────────────────────┬───────────────────────┘
                                       │
                     ┌─────────────────┴─────────────────┐
                     ▼                                   ▼
          Confidence >= 0.85                  Confidence < 0.85 OR
        [Instant Deterministic Action]       Multimodal Book Photo
        - Sub-400ms WhatsApp Reply          [SYSTEM 2: AWS Strands Agent]
        - Direct DynamoDB GSI Match         - Bedrock Nova Pro Vision
        - 48-Hour Reservation Lock          - Multi-Turn Autonomous Reasoning
                                            - Zod-Validated Tool Suites

import { TypeSafeClient, choice, noul } from '@typesafe-ai/sdk';

export interface SystemOneDecision {
  intent:
    | 'offer'
    | 'demand'
    | 'catalog'
    | 'demand_board'
    | 'offer_inquiry'
    | 'demand_inquiry'
    | 'confirm_handover'
    | 'parent_activity'
    | 'contact_inquiry'
    | 'other_grades'
    | 'greeting';
  lang: 'en' | 'fr';
  isGradeMissing: boolean;
  confidence: number;
  probabilities: Record<string, number>;
  model: string;
  latencyMs: number;
  source: 'jev' | 'fallback';
}

export async function triageParentRequestWithJev(
  sanitizedMessageText: string
): Promise<SystemOneDecision> {
  const startTime = Date.now();
  const text = (sanitizedMessageText || '').trim();

  const client = getTypeSafeClient();
  if (client && text.length > 0) {
    try {
      const response = await client.systemOne({
        state: { parent_message: text },
        questions: {
          intent: choice('What is the primary action or objective of the parent?', {
            offer: 'Parent is listing, selling, donating, or has a specific named textbook/subject (e.g. "I have Year 5 Maths", "J\'ai Terminale C Physique").',
            demand: 'Parent is explicitly looking for, wanting to buy, or needing a specific named textbook/subject (e.g. "Looking for Year 8 Maths", "Need Physics", "Je cherche maths 5ème").',
            catalog: 'Parent is asking to view, browse, or see available books in stock (e.g. "catalog", "catalogue", "Year 3 books").',
            demand_board: 'Parent is asking to view the community wishlist or books other parents need (e.g. "wishlist", "demandes").',
            offer_inquiry: 'Parent states vaguely that they have books to offer without naming any specific subject or title.',
            demand_inquiry: 'Parent states vaguely that they need books without naming any specific subject or title.',
            confirm_handover: 'Parent is confirming that an exchange took place, book was sold, or delivered (e.g. "sold", "vendu", "remis").',
            parent_activity: 'Parent wants to see their personal listings, sales, or reservations (e.g. "my books", "mes livres").',
            contact_inquiry: 'Parent is asking for contact info or phone number of the matched parent/seller.',
            other_grades: 'Parent wants to browse overflow secondary grades outside the primary list.',
            greeting: 'Chit-chat, greeting ("hi", "bonjour"), or asking for help or instructions.',
          }),
          language: choice('What language is the parent using?', {
            en: 'English language or English curriculum levels (Year 5, etc.)',
            fr: 'French language or Francophone curriculum levels (6ème, 3ème, Terminale, etc.)',
          }),
          is_grade_missing: noul(
            'Does the message omit the specific curriculum grade, school year, or class level (e.g. Year 5, 3ème, Class 6)?'
          ),
        },
      });

      const latencyMs = Date.now() - startTime;
      metrics.emit('JevLatency', latencyMs, { unit: 'Milliseconds' });
      metrics.emit('JevConfidence', Math.round(response.answers.intent.confidence * 100), {
        unit: 'Percent',
        dimensions: { intent: response.answers.intent.choice },
      });

      return {
        intent: response.answers.intent.choice as SystemOneDecision['intent'],
        lang: response.answers.language.choice as 'en' | 'fr',
        isGradeMissing: response.answers.is_grade_missing.noul > 0.65,
        confidence: response.answers.intent.confidence,
        probabilities: response.answers.intent.probabilities as Record<string, number>,
        model: response.model,
        latencyMs,
        source: 'jev',
      };
    } catch (err: any) {
      metrics.emit('JevFailure', 1, { unit: 'Count' });
      console.warn('[Jev System One] API error, falling back to deterministic fast-path:', err);
    }
  }

  // Graceful deterministic fallback ensures zero downtime
  return executeDeterministicFallback(text, startTime);
}

export const booksAgent = new Agent(scope, 'books-agent', {
  model: { deployed: BedrockModels.BALANCED },
  systemPrompt: `You are Relay, an autonomous AI assistant for parent school textbook exchanges...`,
  tools: (tool) => ({
    searchInventory: tool(...),
    createDemand: tool(...),
    registerBookOffer: tool(...),
    triageParentIntent: tool({
      description: 'Fast System 1 triage of parent message intent, language, and grade completeness using Jev',
      parameters: z.object({
        message: z.string().describe("The parent's message text"),
      }),
      handler: async ({ input }) => {
        const decision = await triageParentRequestWithJev(input.message);
        return {
          intent: decision.intent,
          lang: decision.lang,
          isGradeMissing: decision.isGradeMissing,
          confidence: decision.confidence,
          model: decision.model,
          latencyMs: decision.latencyMs,
          source: decision.source,
        };
      },
    }),
  }),
});