Smart Technology · Government · Easy Access Development Phase

SARAL — System for Algorithmic Fairness in Responsible Access to Livelihoods

A simple way to find and apply for government help using text messages, with clear guidance in your language and protecting your privacy.

AI in v1: lightweight, on-server

Lead Researcher & Developer

2025 – Ongoing

Building first version; pilots targeted for December 2025

SMS/USSD first Vernacular UX AI triage Eligibility hints Bias monitors Minimal PII

Python · Flask · Twilio · MySQL · React

Want methods and early findings?

Many people who qualify for government help never get it, not because they don't need it, but because the process is confusing and hard to navigate.

Information is scattered: Rules and required documents are spread across different websites and paper forms.
Technology barriers: Not everyone has internet access, smartphones, or knows how to use complex websites.
Unclear process: It's hard to track what happens after you apply and what to do next.
Confusing language: Government forms use complicated words that are hard to understand.

Start where people are comfortable; use smart technology to assist, not replace human judgment.

Works on any phone: Start with simple text messages, then upgrade to apps when you're ready.
Simple questions: Easy conversation that helps you find the right programs and documents you need.
Clear explanations: Smart suggestions in your local language that make sense.
You control your data: We only collect what's absolutely necessary and you can download your records anytime.
AI that assists, not decides: Small models help route intents and suggest eligibility hints; a rules engine and humans handle decisions.

SARAL

Text message conversation that narrows to relevant programs in about a minute.

SARAL

Question flow maps answers to programs and a minimal document checklist. Guidance is explanatory—never decisioning.

Simple words, easy steps, and clear guidance every step of the way.

Plain language: Short, clear phrases in your language; pictures only when they help.
Smart document lists: Rules + AI hints surface acceptable alternatives (e.g., any one of PAN / Voter ID / DL) with clear rationale.
Save your progress: Get a short code to continue later, even without internet.
Show what you need: Only see what's important right now; get more details when you want them.

Small, on-server models that assist routing and explanations, with strict confidence thresholds and conservative fallbacks.

See Privacy & Data

Tasks: SMS/USSD intent routing; eligibility hinting (maps answers to likely programs); anomaly flags for review.
Models: Lightweight text classifiers and template-guided generation (no end-user free-text stored).
Data: Hand-crafted rules; small, curated examples; synthetic variants for language coverage.
Confidence & fallbacks: Below threshold → rules-only guidance; operator escalation for edge cases.
Fairness checks: Simple monitors track suggestion rates by language and region; alerts trigger manual audit.
Limits: Hints are non-binding; official eligibility remains with the program authority.

With version 1 on a verge of completion, field testing begins in December 2025 to evaluate usability, fairness, and real-world reliability.

Run field tests (Dec 2025 – Feb 2026): Collaborate with partner communities to observe how citizens interact with SARAL in real settings.
Collect feedback & audit outcomes: Document performance, edge cases, and user experiences to identify barriers or biases.
Refine for version 2: Rework the intake logic, interface, and fairness checks based on testing data and feedback.
Design transparency tools: Build visual summaries and dashboards to help users and auditors trace how decisions are made.
Publish findings: Release a short report and methodology notes in early 2026 to share lessons and invite collaboration.

Version 2 development will begin mid-2026, incorporating findings from the field.

Your privacy and security are built into every part of the system from the ground up.

Minimal personal information: We only collect what's absolutely necessary to help you find programs; most information stays on your device.
You control your data: Download a complete record of your questions and answers anytime.
Strong security: All connections are encrypted; security keys are regularly updated; everything is verified for later reference.
Regular reviews: Privacy and accessibility experts review the system before each new test.
AI privacy: Models run on our servers; prompts are minimised; no raw free-text or personal identifiers are retained beyond short-term processing logs.

Methods we're using now and the metrics we'll track during pilot testing.

Results reflect early internal runs and will be updated after field testing.