Sauti Yo

The problem

Someone facing domestic violence, child abuse, an unfair dismissal or a land dispute may not know the law is on their side, or what to do about it. Legal information is hard to find and written in legal language. And in Uganda, a smartphone, mobile data or a reliable connection can't be assumed.

Sauti Yo, built for the Legal & Policy Advocacy track, takes a person from their situation to verified rights information, practical next steps and a trusted organisation, on whatever phone they have.

How it works

One shared engine sits behind every channel, so guidance is the same whether someone dials a USSD code, sends a text, calls or uses the web:

1. Recognise: describe what's happening in ordinary words.
2. Understand: get the matching verified rights information in plain language.
3. Act: see practical next steps.
4. Connect: reach a verified support organisation, with consent.
5. Be heard: optionally say what made help hard to get. Partner organisations see the answers only in aggregate.

What I built

Keeping AI away from what matters

AI makes replies warmer and handles messages the keyword matcher can't place. But in a crisis, a wrong word or a wrong phone number can put someone at risk. So the rules are strict:

Rewording that can't change a phone number

When AI rewords a verified reply, the numbers in it must come back exactly as they went in. If one is dropped, changed or invented, the rewording is thrown away:

phone_numbers = re.findall(r"\d{3,}", template_text)
# ... ask the model to reword, with a 5-second timeout ...
if set(re.findall(r"\d{3,}", reworded)) != set(phone_numbers):
    logger.warning("Reworded reply altered a phone number, discarding")
    return None   # caller sends the verified original

Where it stands