AWS Builder Center
Agents for Humans: Building Joruri Shohay — An Emergency Helper Agent for Bangladesh 2nd Part

Agents for Humans: Building Joruri Shohay — An Emergency Helper Agent for Bangladesh 2nd Part

The follow-up to my first post: what I actually built for the Agents for Humans hackathon, the bugs I hit along the way, and what I'd tell someone starting from zero AWS experience.

Series: #AgentsofFootball (3 articles)

  1. 3
    Agents for Humans: Building Joruri Shohay — An Emergency Helper Agent for Bangladesh 2nd Part This article
In my first post, I said I'd follow up with what I actually built once it was submitted. This is that post.

What I built

An emergency helper agent for Bangladesh, called জরুরি সহায় (Joruri Shohay) — Bengali for "Emergency Help."
It's a bilingual chat agent that answers three questions in one calm reply:
  • Who do I call right now? — real verified emergency numbers (999, 1090, 109)
  • What do I do while waiting? — standard first-aid steps for the specific situation
  • Where do I go? — nearby hospitals filtered by emergency type and budget, plus blood banks
You type in Bangla or English, and it replies in the same language. It only handles medical and transport emergencies — ask it about restaurants or the weather and it politely says that's outside what it does.

The rule I cared about most

The agent never invents information.
Every phone number, hospital name, and coordinate it states comes directly from a tool result. If it doesn't know something, it says so honestly and points to 999 or 1090 — it doesn't guess.
That matters because this is a safety-critical tool. In a real emergency, a confidently-stated wrong hospital name could cost someone their life. "I don't know, here's the real number to call" is a better answer than a plausible-sounding guess.

How it works

The agent is built with the Strands Agents SDK and runs on Claude Sonnet 4.6 through Amazon Bedrock.
It uses a multi-agent architecture. A primary agent handles the conversation, triage, and safety rules. When it needs hospital information, it delegates that to a separate Hospital Finder specialist — a different Strands agent with its own system prompt and its own tools, running on a smaller Claude model. The primary agent has no direct access to hospital lookup, so every hospital search goes through delegation. That keeps each agent focused on one job.
Seven tools do the work: converting area names into real coordinates, finding hospitals by type and budget, finding blood banks, returning the emergency number, giving first-aid steps, logging user corrections for review, and drafting a message the user can copy and send to family.
Safety is enforced in two layers:
  • Two deterministic hooks that run in code — one caps tool calls per turn, one ensures blood-bank lookup has real coordinates
  • Two LLM-based second opinions that review every reply before it ships — one checks the reply against the safety rules, one keeps the agent inside emergency help only
The whole thing streams responses via Server-Sent Events and keeps per-chat memory using Strands'  FileSessionManager .

The two problems that actually cost me the most time

Problem one: there's no good data source for Bangladeshi medical services.
No API for hospital locations, specialties, budget tiers, or blood bank details. I built the database by hand, checking names and specialties one by one.
This is exactly why the agent is forbidden from stating anything that didn't come from a tool result. After building the dataset by hand, I knew exactly how easy it would be for a model to "helpfully" fill in a plausible hospital for an area that isn't covered yet.
Problem two: a strange Strands bug I still don't fully understand.
At one point, Strands suddenly couldn't recognize or invoke the Claude Sonnet 4.6 model ID at all. Every call failed. No code had changed on my end.
Then I opened the AWS Bedrock Playground in a browser, typed "hi," got a response there — and Strands started working again on the very next attempt.
I don't know the exact mechanism. My best guess is that some provisioning or model-access step on Bedrock's side only fully activates after a first direct call through the AWS console. I'm writing it down here in case another team runs into the same thing.

What I'd tell someone starting from zero

Two things.
One: in a safety-critical agent, most of the actual safety doesn't come from the model. It comes from how narrowly you scope the tools, and how explicitly the system prompt forbids the model from filling gaps with its own training knowledge. Models are capable, but they're also too willing to be helpful. The prompt and the guardrails have to keep it inside a fence.
Two: splitting work across two agents with different scopes produces a cleaner, faster system than one agent trying to do everything. The primary agent stays focused on the conversation. The specialist stays focused on hospital lookup. Each one has less to get wrong.

Honest limits

This project still has real gaps, and I'd rather say so than pretend it's finished:
  • Hospital and blood bank coverage only includes Dhaka, Chittagong, Sylhet, and Khulna
  • Distances are straight-line, not real road distances
  • There's no WhatsApp interface yet — and realistically, nobody opens a website in an emergency
  • Live blood stock isn't tracked anywhere, because no reliable source exists for it
The next version would focus on WhatsApp first, then expand hospital coverage to all 64 districts.

Series: #AgentsofFootball (3 articles)

  1. 3
    Agents for Humans: Building Joruri Shohay — An Emergency Helper Agent for Bangladesh 2nd Part This article
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article