The problem
Indian smallholder farmers don't lose crops because advice
doesn't exist. They lose them because advice arrives after
the spray window closes.
What I built
AgriNexus AI — a WhatsApp advisor that follows up until the
farmer confirms "हो गया" (done).
The architecture (the bit devs will care about)
Three decisions worth sharing:
1. EventBridge Scheduler > Step Functions Wait States
Keeping a state machine open for 48h is expensive at scale.
Step Functions completes in seconds; EventBridge Scheduler
creates one-shot targets at T+24h / T+48h. DynamoDB Streams
→ ResponseDetector Lambda cancels schedules on "done."
Linear cost scaling instead of exponential state transitions.
2. S3 Vectors > OpenSearch Serverless
OpenSearch's always-on OCU costs dominated early bills
regardless of query volume. S3 Vectors eliminated that.
Modeled cost: ~$0.54/farmer/year at 10K scale.
3. Bedrock RAG with visible citations
Retrieve-and-Generate API + Claude. Knowledge base: ICAR,
FAO, NFSM PDFs. Every response has a source link visible to
the farmer. Trust needs traceability.
The stack
API Gateway + WAF → Lambda → SQS FIFO → Bedrock (Claude +
S3 Vectors KB) → Transcribe → Polly → EventBridge Scheduler
→ DynamoDB Streams
Full article with architecture diagrams and ADRs: [article link]
Repo: https://github.com/prasadt1/agrinexus-ai
One ask
This is an AWS Builder AIdeas 2025 finalist. Community voting runs April 18–23 PT. If the architecture or the mission resonates:
click the 👍 like button at the top of this AWS Builder article here 👉 https://builder.aws.com/content/3C8hBRTcsRuQrHzE3Pq243yhXTF/aideas-finalist-agrinexus-ai
(One-time ~30-sec sign-up with Amazon Builder.)
Happy to answer technical questions in the comments.