
Weekend Creative Agent Challenge: “SecretarIA”
AI Receptionist Agent for WhatsApp with Amazon Bedrock.
Vision: The Problem It Solves
Small businesses lose customers when they don't respond outside business hours. In Brazil alone, WhatsApp has over 197 million active users and is the #1 channel for business communication. Clinics, barbershops, pet shops, and restaurants receive dozens of messages daily — but most operate only during business hours.
When a customer sends a message at 8 PM asking to schedule an appointment, nobody answers. Research shows that 78% of consumers buy from whoever responds first. Every unanswered message is a lost customer.
SecretarIA is an autonomous AI agent that acts as a receptionist via WhatsApp — not just a chatbot that answers questions, but an agent that makes decisions: it retains undecided customers, redirects sensitive topics, and escalates to a human when it detects frustration or complex requests. It responds 24/7, knows the business schedule in real time, enforces compliance rules (never gives diagnoses or prices), and costs less than $10/month. Built entirely with AWS services and presented at the AWS Community Day Southeast Brazil 2026.
How It Works
- Customer sends a WhatsApp message
- Backend receives it and applies a 3-second debounce (consolidates rapid messages)
- Message is queued in Amazon SQS (async processing)
- AWS Lambda processes: checks idempotency → loads context → calls Amazon Bedrock → returns response
- Customer receives a natural, helpful reply in 2-5 seconds
The agent doesn't just answer questions — it behaves as an autonomous agent that makes real-time decisions:
- Retention: If someone is undecided, it proactively explains benefits and offers a no-commitment evaluation
- Redirection: If someone asks for prices or diagnoses, it steers the conversation toward scheduling
- Escalation: If it detects frustration or a request beyond its scope, it transfers to a human
- Context-aware: It remembers previous interactions and adapts its tone accordingly
It never says "I can't help you" — it always offers an alternative path.
How I Built It
Process and Key Decisions
The project started from a real need: dental and medical clinics in Brazil receive 200+ WhatsApp messages/day and can't keep up — especially outside business hours. I went through several iterations:
1. Database choice: DynamoDB over RDS
- Pay-per-request ($0 when idle) vs. RDS running 24/7 (~$15-30/month)
- Native TTL for automatic data expiration (privacy compliance)
- No VPC needed = faster cold starts
2. Lambda outside VPC
- Cold start ~600ms (vs. 2-5s inside VPC with ENI attachment)
- $0 extra cost (no VPC Endpoints or NAT Gateway needed)
- DynamoDB, Bedrock, S3 all accessible via IAM — no private subnet required
3. Debounce strategy
- Users often send 3-4 messages in sequence ("Hi" → "I want to schedule" → "orthodontics")
- 3-second debounce consolidates into one Bedrock call
- Saves 40-60% on API costs during peak hours
4. Model selection: Amazon Nova Micro
- WhatsApp responses are short (2-3 lines) — no need for a large model
- ~R$ 0.0004 per message (~$0.00008 USD)
- 1-2 second latency — perfect for chat
Challenges
- Prompt engineering for safety: Getting the AI to never provide diagnoses or prices while still being warm and helpful required extensive testing
- Idempotency: SQS can deliver messages more than once; DynamoDB conditional writes prevent duplicate responses
- Context management: Sending full conversation history is expensive; limiting to last 6 messages keeps costs low while maintaining conversational coherence
AWS Services Used
Architecture Overview
Customer (WhatsApp) │ ▼┌──────────────────────────────────┐│ Backend (NestJS on VPS) ││ Receives message → Debounce 3s ││ Queues to SQS │└────────────┬─────────────────────┘ ▼┌──────────────────────────────────┐│ Amazon SQS (retry 3x + DLQ) │└────────────┬─────────────────────┘ ▼┌──────────────────────────────────────────────────┐│ AWS Lambda (Python 3.12, 512MB) ││ Idempotency → Context → Bedrock → Response │└───────┬──────────┬──────────┬────────────────────┘ ▼ ▼ ▼ ┌────────┐ ┌────────┐ ┌──────┐ │Bedrock │ │DynamoDB│ │ S3 │ │Nova │ │3 tables│ │audit │ │Micro + │ │+ TTL │ │trail │ │Guards │ │ │ │ │ └────────┘ └────────┘ └──────┘Services Breakdown
Service | Role |
|---|---|
Amazon Bedrock (Nova Micro) | AI inference + Guardrails for content safety |
AWS Lambda | Serverless message processing |
Amazon SQS | Async queue with retry + Dead Letter Queue |
Amazon DynamoDB | Patient context, conversation history, idempotency (all with TTL) |
Amazon S3 | Anonymized audit trail |
AWS Secrets Manager | API keys and credentials |
Amazon CloudWatch | Metrics, alarms, structured logs |
AWS IAM | Least-privilege policies per Lambda |
Amazon Bedrock Guardrails
This is the component that makes the system production-ready:
- Topic filters: Blocks medical diagnoses, pricing, off-topic conversations
- PII protection: Anonymizes emails/phone numbers, blocks credit card numbers
- Prompt injection defense: Detects and blocks manipulation attempts
- Cost benefit: When Guardrails block at pre-processing, the model is never invoked — zero token cost
Infrastructure as Code
Everything is declared in Terraform. A single
terraform apply provisions the entire stack in 3-5 minutes.
What I Learned
1. Guardrails are non-negotiable for production AI
Without them, it's a demo. With them, it's a product. Bedrock Guardrails saved weeks of manual filtering implementation.
2. Serverless + AI is incredibly cost-effective
The entire AWS stack costs ~$2.30/month for 50 messages/day. The expensive part is the VPS for WhatsApp connectivity ($6/month), not the AI.
3. Prompt engineering > model size
Amazon Nova Micro (the cheapest Bedrock model) performs excellently for short, objective responses. A well-crafted prompt with clear rules matters more than throwing a larger model at the problem.
4. Privacy compliance can be architectural
DynamoDB TTL automatically deletes data after the retention period. No cron jobs, no cleanup Lambdas — compliance by design.
5. Debounce is essential for messaging apps
Without it, you pay 3-4x more in API calls during peak hours. A simple 3-second wait saves 40-60% on costs.
Results
Metric | Value |
|---|---|
Response time | 2-5 seconds |
Availability | 24/7 |
Monthly cost | ~$8.50 (AWS + VPS) |
Human escalation rate | <5% |
Cost per message | $0.00008 |
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article