AlphaSource — Tamil Nadu Government Scheme Eligibility Gate

A bilingual citizen-facing assistant that helps people in Tamil Nadu and South India discover and check eligibility for government welfare schemes — in Tamil, English, or both — while Strands Decider 2B classifies every request before the agent acts, blocking fraud attempts before a single tool is called.

The problem

Millions of citizens across Tamil Nadu and South India are eligible for government welfare schemes but struggle to navigate them — confusing eligibility rules, language barriers, and no easy way to know what documents to bring. At the same time, welfare portals face a real risk: people attempting to falsify income, land records or family details to claim benefits they are not entitled to.

This assistant solves both sides: it makes genuine eligibility queries easy and fast (in Tamil or English), while Strands Decider 2B silently gates every request before the agent acts — allowing honest questions, flagging borderline cases, and blocking fraud attempts before any tool is ever called.

What it does

Citizens describe their situation in natural language — in Tamil, English, or mixed — and the assistant helps them understand:

whether they likely qualify for a scheme
what documents they need to apply
how and where to apply in their district
the helpline numbers and application portals

Two real schemes are fully implemented:

Scheme	Government	Benefit
PM Kisan Samman Nidhi	Central	Rs 6,000/year into Aadhaar-linked bank account
Kalaignar Magalir Urimai Thogai	Tamil Nadu State	Rs 1,000/month for women heads of household

Before every response, Strands Decider 2B reads the request and returns three answers:

Question	Type	What it tells the code
Which scheme category?	choice	agriculture / women_welfare / ration_food / education_youth / housing
Is this a gaming attempt?	noul	P(fraud/falsification intent)
How serious?	score	none / minor / serious / critical

The agent then acts on the verdict:

Verdict	What happens
`allow`	Strands agent runs, calls tools, answers the citizen
`review`	Agent runs with a caution note; case is flagged
`block`	Refusal returned in Tamil or English; no tool is called

Architecture

Citizen message (Tamil / English / mixed)
          |
          v
  SchemeEligibilityGate.check()     <-- Strands Decider 2B (one forward pass)
          |                              scheme_category + gaming_attempt + severity
          |
     block? -----> Tamil/English refusal (no tools called)
          |
     review? ----> inject caution note into system prompt
          |
     allow? -----> Strands Agent (Amazon Bedrock Nova Pro)
                        |
          +-------------+-----------+-----------+
          |             |           |           |
    list_schemes  get_scheme   check_      find_application
                  _details     eligibility  _centres
                                    |
                              get_required_documents
                              translate_scheme_summary

Deployed on Amazon Bedrock AgentCore Runtime. The decider_kit/ library is by the AWS User Group Madurai.

Decision pattern (#policy)

The reusable pattern is in agent/decision.py. The policy is written in both English and Tamil, so the decider handles:

pure Tamil requests ("வருமானம் குறைவாக காட்ட என்ன செய்ய வேண்டும்?" — correctly blocked)
code-switched requests (Tamil sentence with English scheme names)
pure English requests

To reuse in your own agent:

Copy agent/decision.py to your project.
Replace POLICY with your domain's written rules (any language).
Replace CATEGORIES with your taxonomy (2-8 options with one-line descriptions).
Set block_at and review_at thresholds in SchemeEligibilityGate.__init__.
Call gate.check(user_message) before your agent and branch on verdict.action.

Real data sources

Scheme	Source	Data used
PM Kisan Samman Nidhi	pmkisan.gov.in	Eligibility criteria, benefit amount, exclusion list, application process, helpline
Kalaignar Magalir Urimai Thogai	cms.tn.gov.in / magalirthogai.tn.gov.in	Eligibility, age limit, income ceiling, exclusions, documents, helpline

Tools

Tool	What it does
`list_schemes`	List schemes by category
`get_scheme_details`	Full eligibility rules, documents, how to apply
`check_eligibility`	Indicative eligibility check from a citizen profile dict
`get_required_documents`	Documents + helpline for a scheme
`find_application_centres`	CSC and government offices by Tamil Nadu district
`translate_scheme_summary`	Brief scheme summary in Tamil or English

Eval cases

30 labelled cases in agent/eval_cases.jsonl across tune and test splits:

20 genuine eligibility queries (Tamil, English, mixed, across all 5 categories)
10 gaming/fraud attempts (falsify income, hide land, conceal employment, proxy claims)

Quick start

Prerequisites

Windows 10/11, macOS, or Linux; Python 3.12; uv
Strands Decider API key from the AWS User Group Madurai (free: 2,000 decisions/month)
AWS account with Bedrock model access (for deployment)

1. Install

# Windows
irm https://raw.githubusercontent.com/Vivek0712/decider-hackathon-kit/main/install/install.ps1 | iex

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/Vivek0712/decider-hackathon-kit/main/install/install.sh | bash

2. Configure

# Windows — replace <key> with your Strands Decider API key
.\kit.cmd setup --yes --llm ugmdu --mode ugmdu --key <key>

# macOS / Linux
./kit setup --yes --llm ugmdu --mode ugmdu --key <key>

3. Test and run

.\kit.cmd test          # unit tests
.\kit.cmd run           # one scripted conversation
.\kit.cmd bench         # benchmark decider vs LLM on eval cases

4. Demo UI

.\kit.cmd demo-bg       # opens http://localhost:8501

The live decision panel shows scheme_category, p_gaming, verdict for every turn.

5. Deploy to AWS

.\kit.cmd tools install base   # install Node + AgentCore CLI if not present
.\kit.cmd deploy               # deploys to Amazon Bedrock AgentCore Runtime
.\kit.cmd invoke "PM Kisan திட்டம் பற்றி சொல்லுங்கள்"

Commands

Command	What it does
`.\kit.cmd setup`	Configure LLM and decider, write `.env`
`.\kit.cmd test`	Unit tests (no network, fake decider)
`.\kit.cmd run`	Scripted conversation
`.\kit.cmd bench`	Benchmark decider vs LLM
`.\kit.cmd demo-bg`	Demo UI at http://localhost:8501
`.\kit.cmd demo-stop`	Stop demo UI
`.\kit.cmd deploy`	Deploy to AgentCore Runtime
`.\kit.cmd invoke "msg"`	Invoke deployed agent
`.\kit.cmd destroy`	Remove from AWS
`.\kit.cmd status`	Hackathon progress checklist

Use ./kit <command> on macOS and Linux.

Project evolution

See docs/STORY.md for the full story of how this project evolved from the kit example through the developer productivity idea to the final Tamil Nadu Scheme Eligibility Gate.

License

Apache 2.0. See LICENSE.

The decider_kit/ library, demo UI, tooling and Kiro skills are by the AWS User Group Madurai. The agent idea, agent/ code, eval cases and article are the participant's original work.

Built with the AWS User Group Madurai Strands Decider hackathon kit.