-
-
Architecture: End User Messaging SMS and Amazon SES, SNS, Lambda, Bedrock agent, DynamoDB, Step Functions, EventBridge, API Gateway
-
Whanau Hub SMS Attendance Agent on AWS, full architecture diagram with SES staff email and nightly digest
-
Staff console: the Hub's live programme timetable, participants, and the participant phone
-
Conversation thread with tool-call chips, and an escalation emailed to staff via Amazon SES
-
Staff reschedule a session; every enrolled participant is notified by SMS
Inspiration
Kura Cares Charity runs the Whānau Hub (https://www.whanauhub.co.nz) from Takanini in South Auckland: free career, financial literacy, personal development and whānau programmes. The Hub's public page describes "100s whānau supported" across "6 specialist programmes".
The problem is attendance. A lot of participants do not arrive at sessions, so a staff member phones people one by one to remind them. Schedule changes slip through in both directions. Phone calls do not scale or leave a record, and one-way reminders miss the reply, which is where the useful information is.
What it does
The agent sends session reminders by SMS 24 hours and 2 hours before each session. Participants reply in their own words, and an AI agent reads the reply and acts on it:
- "Yes I'll be there" confirms attendance.
- "Can't make tonight, is there another session this week?" looks up alternative sessions in the same programme, offers one, and reschedules when the participant agrees.
- A message that signals distress, a complaint, or a question about eligibility or money is not answered by the model. It is escalated to a staff member with a summary and an urgency level, and staff get an email straight away.
- STOP opts the participant out, and nothing is sent to them again.
Staff use a web console showing sessions, participants, every SMS thread with the agent's tool calls inline, and the escalation queue. When staff move a session, the console texts every enrolled participant. Each evening staff get an email digest of tomorrow's sessions, today's replies and open escalations.
How we built it
Everything runs in ap-southeast-2, deployed with CloudFormation.
- AWS End User Messaging SMS sends every message through the sender ID
EASYCODER, usingSendTextMessageon the Pinpoint SMS and Voice v2 client (boto3.client("pinpoint-sms-voice-v2").send_text_message). It is called from thesend_reminder,inbound_agentandapiLambda functions. - Amazon EventBridge Scheduler starts an AWS Step Functions state machine,
ReminderWorkflow, every 15 minutes. Afind_dueLambda returns enrolments due a reminder and a Map state runssend_reminderfor each one. - Inbound SMS arrive on an Amazon SNS topic,
InboundSms, in the End User Messaging two-way payload format. Theinbound_agentLambda loads the participant and their last 10 messages, then runs an Amazon Bedrock Converse API tool-use loop with Claude Haiku 4.5 (up to 6 turns). The tools areget_my_sessions,confirm_attendance,cancel_attendance,get_alternative_sessions,reschedule,escalate_to_staffandopt_out. - Amazon SES (
sesv2send_email) emails staff each escalation immediately, plus a nightly digest from a scheduleddaily_digestLambda. - Amazon DynamoDB holds all state in one table,
WhanauHubComms. - Amazon API Gateway (HTTP API) and a Lambda function serve the staff console and its JSON API.
- Amazon CloudWatch Logs holds structured logs.
Design choices
- Safe by default. Every send is dry run unless a person switches on "send real SMS". The agent still runs Bedrock, Step Functions, SNS and DynamoDB for real, so the full behaviour is visible before a single message is billed. The sender also refuses any number outside an allowlist.
- Inbound built on the production path. Replies arrive through the same SNS topic a two-way number publishes to. The console's participant phone drives that path today, so the agent is exercised exactly as it will run when the NZ two-way number is attached.
- Decoupled from the Hub. The agent keeps its own attendance record, seeded from the Hub's live programme catalogue, so it works alongside the Hub today and plugs into it through a write-back API without redesign.
Accomplishments that we're proud of
A real SMS from EASYCODER lands on a real phone, a participant answers in their own words, and the record changes without anyone at Kura Cares picking up the phone. The agent confirms, reschedules against the real timetable, respects STOP, and knows exactly when to hand a person to a person, with staff emailed in the same minute. Every reminder reminds whānau that turning up earns points toward their next voucher, which is how the Hub already rewards attendance. The whole thing was designed, built, deployed and tested in one evening on AWS End User Messaging, Amazon SES and Amazon Bedrock.
What we learned
Tool use is what makes this safe. Grounding every answer in a DynamoDB lookup, and giving the model an explicit escalation tool, matters more than the prompt.
Where this goes
Every session at Kura Cares gets reminded, every reply gets understood, every change gets recorded, and staff time goes back to the people who need a person. The same agent serves all six programmes today and any community provider tomorrow: add a programme and a roster and it runs. From here it grows into the channels whānau already use, WhatsApp through End User Messaging Social and RCS as NZ carriers light it up, writes attendance straight back into the Hub, moves the agent loop onto Bedrock AgentCore Runtime, and ships as an AWS Marketplace offering so other charities and training providers can switch it on.
Submission details: ACE opportunity ID O30500895, campaign code "AWS CDS Agentic AI Hackathon -Sept. 2026".
Built With
- api-gateway
- bedrock
- claude
- cloudformation
- dynamodb
- eventbridge
- lambda
- python
- ses
- sms
- sns
- step-functions