Relay. Autonomous School Book Marketplace & Matchmaker

Relay is an autonomous WhatsApp AI agent that eliminates school group-chat book spam by turning casual parent texts and cover photos into instant, proactive textbook matches.

Comment

Youtube Video link is refusing to update.So here it is

https://youtu.be/vtL6SHK-MmQ?si=JBmVCtY8l6y0mgbp

Inspiration

Every August, over 500 families in our school community face the exact same quiet anxiety: back-to-school textbook shopping.

For most of the year, our school WhatsApp group is a vital lifeline. But from August through December, it collapses into utter chaos:

  • Parents whose children just finished Year 4 post lists to sell old books and fund incoming Year 5 curriculum costs.
  • Concurrently, new parents joining Year 4 post frantically trying to buy those exact books second-hand.
  • The Ephemeral Scroll Trap: WhatsApp is a fast-moving linear stream. A listing posted at 9:00 AM vanishes off-screen by 9:20 AM. Desperate to stay visible, parents repost the same lists repeatedly.
  • Communication Paralysis: Searching "Books" in our school chat surfaced over 53 disjointed message threads across Year 1 through Year 13 textbooks, completely displacing official school announcements.

School WhatsApp Group Chat Book Chaos Figure 1: Real-world screenshots from our school community WhatsApp chat. Notice repeated reposts, requests to message in private ("inbox please"), and the total displacement of official school discussions.

Previous attempts by the PTA to organize exchanges with Google Forms, web portals, or mobile apps failed completely, abandoned within 48 hours because parents felt like unpaid data-entry clerks and refused to manage another login for a once-a-year chore.

This inspired our core "Agents for Humans" thesis:

"The best interface for humans is no interface at all. Zero new apps. Zero logins. Zero forms."

We built Relay to meet parents where they already communicate—directly inside WhatsApp—working quietly in the background to turn chaotic chatter into an autonomous, proactive circular textbook economy.

What it does

Relay is an autonomous, bilingual AI agent operating directly inside WhatsApp that coordinates peer-to-peer textbook recycling:

  1. Natural Language & Compound Offer Parsing: Parents text casual messages (e.g., "I have Year 5 Cambridge Maths and Year 6 French textbook for $15"). Relay’s reasoning engine decomposes compound multi-intent messages and catalogs each item in under 800ms.
  2. Multimodal Cover Scanning: Busy parents don't need to type anything. Snapping a photo of a textbook cover triggers Amazon Bedrock Nova Pro Multimodal Vision, which extracts the exact subject, publisher, curriculum grade, and condition automatically.
  3. Proactive Wishlist Matching: When a parent asks for a book that isn't currently listed, Relay records it on the demand board. The instant another parent offers or scans that textbook weeks later, Relay autonomously pairs them.
  4. Autonomous 48-Hour Escrow Reservations: To eliminate school gate race conditions and double-booking, Relay places an atomic 48-hour reservation hold on matched books, issues a mutual 4-digit verification PIN (#7721), and shares safe pickup instructions.
  5. Year-Grouped Demand Board & Live Storefront: Parents can type "demand board" or "catalog" to view community needs organized cleanly by school year (*Year 5*, *Year 10*, *Terminale*), or browse the public web storefront hosted on Amazon CloudFront.
  6. Bilingual Native Intelligence: Operates fluently in both French and English, automatically detecting language and adapting responses.
  7. In-Chat Developer Sandbox: Reviewers and developers can text #SANDBOX ON followed by #SEED directly on WhatsApp to test simulated Cameroon/French curricula and state transitions without touching live community records.

How we built it

Relay is architected on AWS Blocks as a 100% serverless, event-driven system on AWS:

Relay End-to-End Architecture Topology Figure 2: Relay Simplified End-to-End System Architecture. Illustrates WhatsApp parent interaction, Security & Boundary Layer, AWS Blocks foundation, and the autonomous Strands Agents / Amazon Bedrock runtime.

  • Agentic Reasoning Layer: Built with the AWS Strands Agents SDK (@aws-blocks/bb-agent), coordinating multi-turn reasoning loops with strictly typed Zod tool schemas.
  • Amazon Bedrock Multi-Model Architecture:
    • Amazon Nova Lite: Sub-400ms intent classification and structured curriculum entity extraction.
    • Amazon Nova Pro: High-resolution multimodal vision analysis for book cover extraction.
    • Claude 3.5 Sonnet: Nuanced bilingual reasoning and complex conversational edge cases.
  • AWS Lambda Durable Functions: Webhook transactions use an idempotent step-execution engine (withDurableExecution), delivering the fault tolerance of distributed sagas with zero Step Functions state-transition costs.
  • Amazon DynamoDB: Single-table on-demand architecture with Global Secondary Indexes (byConcept) for sub-10ms catalog lookups and wishlist matchmaking.
  • Zero-Trust Community Privacy: Dual-stage in-memory PII redactor scrubs phone numbers and street addresses before prompts reach the LLM, backed by AWS KMS Customer Managed Keys ($AES\text{-}256$ envelope encryption) and constant-time HMAC-SHA256 webhook validation (crypto.timingSafeEqual).

Challenges we ran into

  1. The 3-Second Webhook SLA vs. AI Latency: Meta's WhatsApp Cloud API drops connections if an HTTP 200 isn't returned within 3,000ms, triggering duplicate webhook storms. We decoupled ingress from processing using an asynchronous, idempotent durable execution pattern that acknowledges the webhook in under 300ms while executing Bedrock agentic steps safely in the background.
  2. Preventing AI Hallucinations in Marketplace Matches: LLMs can easily conflate similar subjects (e.g., matching Year 5 Maths with Year 10 Maths). We implemented hybrid determinism: common queries (catalog, my books, demand board) bypass the LLM entirely via sub-50ms regex fast-paths and DynamoDB GSI queries, reserving Bedrock for unstructured natural language and multimodal scans.
  3. School Gate Race Conditions: When a popular textbook was posted, multiple parents rushed to claim it simultaneously. We engineered atomic 48-hour conditional reservation locks in DynamoDB with mutual 4-digit verification PINs to guarantee that only one buyer holds an item at a time.
  4. Zero-Risk Testing for Hackathon Reviewers: We built an in-chat sandbox engine (#SANDBOX ON) that isolates session state and intercepts mock phone dispatches so judges can test full flows without polling or polluting real community inventory.

Accomplishments that we're proud of

  • 🔇 100% Elimination of Book Spam: Over five full months of repetitive book chatter completely vanished from the main school group chat.
  • 📊 Real-World Marketplace Velocity: Over 100 textbooks cataloged, 26 active parent wishlists, 15 active 48-hour holds, and 47 completed exchanges saving families thousands of dollars.

Relay Production Marketplace Live Dashboard Figure 3: Production Dashboard Live Metrics showing active inventory, pending demands, 48h holds, and completed parent exchanges.

  • 💬 Bilingual Community Praise: Overwhelmingly positive reactions from community parents in French and English:
    • 🇫🇷 "Bonsoir chers parents. Wow c'est superbe, merci d'avoir pris la peine de la faire. 😎"
    • 🇫🇷 "Bonjour chers parents très bonne idée j'aurais grand besoin de lui 🙏"
    • 🇬🇧 "Thanks very much. This makes it quite easy. I will type in what I have there."
    • 🇬🇧 "great job! 👍"

Real Community Parent Praise and Feedback for Relay Figure 4: Authentic feedback and testimonials from school community parents in English and French.

  • 💰 Sustainable Micro-Cost Serverless Architecture: The entire AWS cloud infrastructure runs for ~$2.65 / month:

$$\text{Monthly Operating Cost} \approx \$2.65 \quad (\text{Free Tier Optimized})$$

$$\text{Community Financial ROI} = \frac{\text{Direct Family Savings}}{\text{Annual Cloud Cost}} = \frac{\$54,000}{\$13.25} \approx \mathbf{4,075\times}$$

What we learned

  1. The Best Interface is No Interface: When designing agents for humans, never build a custom app when an established messaging channel exists. Meeting parents in WhatsApp eliminated 100% of onboarding friction.
  2. Hybrid Determinism Trumps Pure Generative AI: Routine commands and schema lookups should never touch an LLM. Pairing deterministic code paths with selective agentic models yields sub-50ms latency, zero hallucinations, and negligible cloud costs.
  3. Durable Idempotency is Non-Negotiable: When an AI agent mutates real-world state (reserving books, releasing escrow holds, dispatching contacts), idempotent step engines like withDurableExecution are essential to survive dropped webhooks and network retries.
  4. Community Trust Requires Zero-Trust Privacy: Parents will only adopt an AI assistant if their private phone numbers and family data are protected. In-memory PII redactors and KMS envelope encryption are mandatory baselines.

What's next for Relay. Autonomous School Book Marketplace & Matchmaker

  • 🥋 School Uniform & Sports Gear Recycling: Extending the vision extraction model to catalog and match outgrown school blazers, lab coats, and athletic gear.
  • 📈 PTA Curriculum Deficit Alerts: Autonomous batch reporting that alerts school administrators and parent committees to specific grade-level textbook deficits before the semester begins.
  • 🏫 Multi-School District Federation: Enabling neighboring school campuses to cross-list syllabus-compatible textbooks while preserving localized school gate pickup hubs.

Built With

  • agentcore
  • awsblocks
  • awskms
  • awsstrands
  • awswaf
  • dynamodb
  • guardrails
  • lambdadurablefunctions
  • meta
  • s3vectors
  • strands
  • whatsapp
Share this project:

Updates

Submission history