Every year, 37 lakh people in India get bitten by dogs. Rabies is almost always fatal. The vaccine course is almost always effective. And nearly half the people who start it never finish.
Not because they don't want to. Because they forget. Because the clinic is far. Because nobody reminded them.
That's the problem my teammate and I tried to solve this weekend at AWS First Commit.
What we built
Poora Teeka ("complete the vaccine") is a dose-completion and vial-logistics tool for anti-rabies clinics. A nurse registers a bite patient, picks the protocol, and the system takes over: it generates the dose calendar, sends a WhatsApp reminder with a Hindi voice note the day before each dose, detects missed doses automatically, and groups tomorrow's patients into shared-vial time slots so fewer vials get thrown away.
That last part matters more than it sounds. An opened intradermal vial has an 8-hour window. Clinics routinely discard 20% of their vaccine. Our batching feature cut a 7-vial day down to 3 in our demo.
The AWS stack
We ran everything in ap-south-1 (Mumbai) for data residency.
A few architectural choices worth explaining:
EventBridge Scheduler over a cron. A polling cron that checks "which doses are due tomorrow" scales with table size. EventBridge one-time schedules are O(1) per dose. One schedule per reminder, created at course time, fires once, done.
Postgres over DynamoDB. The product's core value is constraints. A vial cannot serve a negative number of doses. A route cannot change mid-course. These are enforced by the database itself with CHECK constraints and FOR UPDATE SKIP LOCKED, not by application code that can be bypassed. We ran a load test: 20 concurrent writes, one vial, zero invariant violations.
Transactional outbox. Every dose creation writes the reminder and the outbox row in the same transaction. If Lambda crashes halfway through, nothing is lost. The poller picks it up on the next 1-minute run.
Full stack: Lambda (Node.js 22), RDS PostgreSQL, EventBridge, SQS with a DLQ, Amazon Polly for Hindi voice notes, S3, Amplify for the Next.js dashboard, Secrets Manager, CloudWatch. All deployed with a single CDK stack.
The thing that surprised us
We integrated TypeSafe AI's Jev model, released literally 5 days before the hackathon, to triage missed doses. It classifies each one as routine, priority, or critical in under 100ms, post-commit, without blocking the sweep. A patient who missed dose 4 of 5 gets flagged critical. A patient who missed dose 1 of 5 gets a standard reminder. One API call, no prompt engineering, typed output straight into our dashboard.
It's exactly the kind of smart if-statement that belongs inside software, not a chatbot.
What we learned
The hardest part wasn't the code. It was the 3 AM moment when we realized our RDS instance was sitting in Sydney while our Lambdas were in Mumbai. 140ms per query, 4 or 5 queries per handler. Fixed it with a CLI command and a fresh instance.
The most satisfying moment was the WhatsApp message arriving on a real phone with a Hindi voice note playing. That's when the project stopped being an engineering exercise and started feeling like something a nurse could actually use.
Loading article