AWS Builder Center
My Journey at AWS  Bharat Build Hackathon

My Journey at AWS Bharat Build Hackathon

I want to share my experience , journey and idea during the Bharat Build Hackathon by AWS and WeMakeDev. The experience and learning was on different level to share, hope you will like it.

Just Learn , Explore and Transform

MY LITTLE JOURNEY IN - Bharat Build Hackathon - By AWS and WeMakeDev

GreenPhoeinx -- Wondering What is this??

So let me introduce with my project title "GreenPhoenix" . Did you liked the name??
Before I start telling what was my idea , I would like to tell u why I chose this name.
 > So ``Phoenix`` is a ``mythical`` bird who is immortal and helps those who are in need and is born from flames of fire !!! 
 > Why Green?? The color Green makes feel good and healthy and joy full of energy!!! 
 > So GREEN + PHONIX ==> Something that helps ones in need and make them feel happy , healthy and joyfull. 
 IF YOU WANT TO KNOW HOW I CAME UP WITH THIS IDEA - FIND IT AT THE END OF THE ARTICLE. 

My Logo for the Project

App Logo


Project Idea:

So I guess you guessed it right - my idea is to build an application which will help the people in crisis, whatever may be the need.

THE PROBLEM:

Every year, many people face eviction, hunger, or sudden loss of income due to several reasons but hit same wall of not getting any support and then crisis , poverty and several consequences.
The problem is not help or schemes does not exits --- Help exists, but finding out which help will actually be beneficial for the crisis is very frustrating. And On the top it --- Aid programs are scattered across federal, state, city, and nonprofit systems, each with its own eligibility rules, its own paperwork, and its own jargon. SO anyone has to go through lots of documentation or rules before finding which one is actually helpful.
For instance, A single mother losing her housing doesn't need a lecture on the difference between SNAP and the Emergency Rental Assistance Program — she needs someone to tell her, quickly and in her own language, what she qualifies for and how to apply.

Overview of What I want to Build:-

Basically what I want to build is something which will help the one in need due to any reason be it disaster, natural calamity, or any household problem, there are lots of NGOs or Government Schemes for them but most of them are not aware of them.
So my attempt was to build an AI assistant or agentic system to help them which will find correct and customised result for them in there own language (support voice system) based on their background and need. Also auto add to custom community and support groups for them so that everyone can share and help each other. And Finally something to detect biasness and in case of dispute or if more detail is needed, caseworkers will handle the situation and for them a dashboard to track people in need and the issue they are passing , and some dashboard for analyst and admin to check the biasness and other control.

My Implementation:

Some of you might have some doubt - the plan looks nice or atleast decent !!! but what about the implementation, Was I able to make it whole? - My Answer is No - not able to do all of these in 3 days!!!
I guess what I actually was able to implement is just 30-40% of the idea I initially thought.
So as I was aware that I was working solo and not possible to complete this on time as some bugs will come during development and deployment - so I limited my focus mainly on these:
  • Make a multi-agent orchestration pipeline to help the users in need from collecting information to matching programs and help in documentation.
  • Every User can use the website and chat in there own language
  • Remove PII(Personal Information) so that user 's data remains safe .
  • A working backend to use the endpoints and respective dashboard in frontend/client for every role - which includes - PublicApplicant, CaseWorker, Analyst, and Admin
  • Something to check for auditing to ensure fairness and biasness .

The Challenges:

So What can be the challenges ??
  • Major Challenge - is how can I make my agent aware of the programs and its detail.
  • So to solve this I was thinking about 2 possibilities - give agent a tool to web search + scrap data from websites, or make a dataset with some real and dunny schemes ( May do some kind of partner-ship to use actual data ) or some kind of knowledge base.
  • MY Approach: - For the Hackathon I choose to make a dataset (Since other option is better but will add to complexity ) and use OpenSearch to search.
  • How to deal with the problem that Common People face ?? (USER's Perspective) :
  • What I was thinking , to find right schemes or programs -- what information I need and what user want from my side - SO what I think is -- Someone to tell - which schemes to apply and what documents do they need based on their situation and background.
  • Also people in need have no intention to login and so decided keep it public, and to prevent misuse implemented - rate limiting.
  • My Approach:- Make 4 agents - Intake (Who will try to understand the situationa and background), Matching Agent (Who will match the programs from dataset they qualify), Document Agent (Who will help them in drafting application and compiling checklist), AuditAgent (Who can separately run test and check for biasness)
  • Other Challenges:
  • What kind of access and roles and permission I should give ?
  • Weather add login/signup for normal Applicant? - No login for applicant but for sesititive roles and implemented rate limiting for users to prevent misuse.
  • What tools should I make for the agents - make it deterministic or non-deterministic
  • What tech stacks to use ?
  • How to handle PII and User Privacy? How to manage Session??

Tech Stacks Used :

  • FastAPI - For creating the api endpoints, and backend routes.
  • AWS Strands: (Agent SDK) - The core of the system: Intake, Matching, Document, and Bias-Audit agents, plus the top-level orchestrator coordinating them as tools.
  • Presidio (from Microsoft) - Open Source library for custom PII removal and recognisers.
  • Cedar — role-based, resource-scoped policy engine gating every tool call (search, eligibility check, document draft) and every FastAPI route, with a dedicated policy for PII vault access and a human-review gate for high-stakes programs
  • Amazon OpenSearch (self-hosted via Docker locally) — dual-purpose as both the BM25 keyword index and the k-NN vector store for hybrid semantic search over aid programs and also backs the audit-log and biasness-audit-log index.
  • Amazon DynamoDB (DynamoDB Local for dev, real DynamoDB in prod) — staff user credentials store (Caseworker/Analyst/Admin) and storing language translations .
  • Amazon Translate — normalizing non-English input to English before it reaches the Intake agent, and translating responses back to that language.
  • BedRock - LLM provider in agents with gemini as a fallback.
  • Vite+React( bun - package manager) and Docker - For the frontend and individual dashboard and containerisation and running some services locally.
  • Nginx + GitHub Actions - For reverse proxy and for CI/CD in EC2
  • Deployment - EC2 t3-large and Elastic IP for static IP4 address and IAM role for EC2 host to use the AWS services.
  • SOME PLANNED BUT NOT USED:-
  • Amazon Comprehend — language sentiment detection.
  • Amazon Polly — text-to-speech output, using Neural/Long-Form voices and SSML for pacing
  • Amazon Transcribe — streaming speech-to-text for voice input
  • Amazon Bedrock Knowledge Bases — proposed for a separate FAQ/process-question index, reusing the same OpenSearch cluster as its vector store
  • Amazon Textract — OCR/document verification for uploaded proof documents

The Architecture Diagram ( Though not accurate)

(Generated by Gemini)
architecture diagram

So overall the architecture and agent loop looks some what like this ... Just in the agent loop the audit agent acts separately.
"""SO Now I have my plan and everything ready --- It was time to build."""

BUILD PHASE:

Now actual journey starts :
  • First initialised both frontend and backend ( For Frontend - used bun as package manger and React+Vite and For Backend - used uv as package manager and FastAPI ) , made a csv file and used docker for opensearch and redis.
  • Then most challenging was to built the agent orchestration layer - need to fix multiple times and adding to the pity 🥲 ( "BedRock was not working and I need to use Gemini in free tier - getting 429 many times" ) - need to add deterministic fallback.
  • Then went on to make api end points , role based access using cedar policies and other routes for separate roles and authentication.
  • Since decided to keep the endpoints pubilic for the end users in need - have to implement rate limiting.
  • Focused on frontend building separate dashboard for different role and only displaying as per the role access.
  • Adding Multilingual support using AWS translate and in frontend using i18 fixing UI used tailwind and lucide-react , framer motion.
  • Refactored for deployment and final touch - making docker file and nginx config and other small task.
  • Finally Deployment in EC2 setup and setting up IAM roles and configs in the VM and allocating Elastic IP.

Challenges During Deployment:-

  • So I was not saving the chat history in DB to ensure privacy. So a challenge was - to ensure history persist across refreshs within the same session - so to do this used - Redis to store session with a TTL of 30 days.
  • For translation converting takes time using AWS translate - so used redis to store the cached so that it's fast - and used dynamo DB to store these since most of the translations are same .
  • The agents were specific to foreign American region as initially my csc file was full of American programs- fixed it included Indian programs and customised for global usecase.
  • The agents loop was too strict to ask straight questions without looking the intent of the user - fixed that corrected system prompt and also ans as per global regions not specific to something.
  • Several other smaller bugs and Issues came during both coding and deployment too.

A Lesson I think was Imp was:

"Use AI coding agents and tool to fast the development not to build the whole application"
So I was using agy cli to fasten - although new with this ( has used before but not for heavy tasks) -- the code which initially looked good performed terribly bad than what was my plan and vision. Good for bug fixing and particular task - but not for building whole application, the 3rd and 4th challenge and several other was specifically due to that.

FINALLY A ICEBREAKER - HOW I GOT THE IDEA :

So recently in Nepal , a huge flood came and lots of people lost their houses , lands and family members and lots of help came lot of funds and government support and NGOs working there. But my question is the people in need really got the help and what about the paper work ?? How will they come to now that help exits and which one they are eligible for?? They are already frustrated.
SO I came with this --- however I feel that whatever I tried to make and what was my vision -- I was not able to fulfil or make something like that in this limited time.
Anyway -- HOPE EVERY ONE STAY WELL AND HEALTHY --
"" At the end, THANKS TO AWS, WeMakeDev and everyone who was a part of this wonderful journey either directly or indirectly. I learned a lot and really a great experience hoping to know a lot more in future. It will definitely a wonderful experience to remember and this will help me a lot in my upcoming journey ahead. ""
stay healthy

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