Roshni

Roshni: light, to illuminate and to make clear.

Inspiration

Roshni came from a very deliberate decision: we did not want to build a generic AI project. Our team is proudly Pakistani, and we wanted to work on something that mattered to the people around us. We chose BISP (the Benazir Income Support Programme), Pakistan's largest social safety net, which gives families living in extreme poverty regular cash stipends. The money is paid directly to the mother of the household, and BISP runs several subprograms on top of it, such as the Taleemi Wazaif education stipend, which pays mothers to keep their children in school, and the Nashonuma program for maternal and child nutrition.

But support existing is not the same as support reaching the people it was built for. Picture a mother in a village in rural Punjab, where fewer than half of rural women in the country can read (around 47.5%, according to the 2023 national census). A message arrives about her child's education stipend. She cannot fully read it, she cannot tell whether it is real or one of the many fake BISP messages circulating, and she does not know what she is meant to do or by when. So the stipend that exists for her family quietly slips away, not because she did not qualify, but because she could not understand the message telling her that she did.

We found the unfairness of this infuriating. The help was already hers. The only things standing between her and it were language and red tape. As high school students, we cannot change government policy. But we could build something practical: a tool that helps a mother read the message, understand it, and get the stipend to her child.

What it does

Roshni is an AI-powered tool for official, BISP-related messages. The user simply pastes in the message they received. After processing it, Roshni explains it in simple, easy-to-understand language: what the message is, whether it applies to the user, what action they need to take, and any deadline mentioned in the message itself. It also gives the message a clear safety verdict, marking it as safe, caution, or a likely scam, and points the user to the official 8171 helpline to verify anything on their own. Roshni can respond in Roman Urdu, Urdu, or English, depending on the user's choice, and has an optional voice output, which we built in for accessibility. Since Pakistan has one of the lowest literacy rates in South Asia, we wanted to make sure users could actually access the information the tool was giving them, so we let them listen to it. Most importantly, Roshni never takes any action on behalf of the user. And whenever a message is unclear, it does not guess: it sends the user to the official 8171 helpline instead.

How we built it

Roshni is built as a web application using FastAPI and the OpenAI API. Our code sends users' messages to OpenAI's GPT-4o model, which is guided by a long and detailed prompt. That prompt contains details of the target audience, the tone the model must adopt, and the language it must avoid to stay easily understood, and it is instructed never to hallucinate dates, links, or actions. Roshni also incorporates a text-to-speech pipeline so that explanations can be listened to aloud.

We spent this week working together by first conducting primary and secondary research on the BISP program, then developing our prototype. We then beta-tested it with synthetic messages modeled on real BISP notices and known scams, incorporated accessibility features, and tweaked our system prompt based on what we found.

Challenges we ran into

One of the largest challenges we faced was making the tool simple to use but genuinely effective. Since we were targeting a very vulnerable section of the population, we wanted to make sure our tool could actually help impoverished people access the support they need. If it were cluttered with features, it would overwhelm the user and become the very problem it was designed to solve. If it were too simple, it would risk losing important meaning. Our project sits in a sensitive space where, if our AI tool were buggy, it could do more harm than good.

Another challenge, related to the first, was focusing tightly on our target audience and working out exactly how to serve their needs. Our tool was inherently text-based. But the issue we wanted to address was that text-based government messages are hard to decipher in a low-literacy environment. That is where our audio and language support came in. We also had to make sure the information itself was pitched right for our audience: we spent a lot of time studying official BISP notices, noticing the common formal lingo, and adding clauses to our system prompt to swap in simpler synonyms.

Accomplishments that we're proud of

First and foremost, our team is proud of picking a well-defined target audience. Coming into this hackathon, we were determined to work on something we were truly passionate about and that we felt could make a real difference. BISP is something most Pakistanis are familiar with, but, at least to our knowledge, no one has focused on building something to address some of the program's limitations. We know Roshni has a long way to go before it could earn government support, but we are proud of the progress we made this week, and it has left us optimistic about the work we can do together. We are also a team with relatively little programming experience, with only one of us having a stronger coding background. We entered USAII hoping to build a UI or a proof-of-concept, and we have exceeded our own expectations for what we could do together.

What we learned

Even if this project does not end up winning, we have learned a lot about how we think about AI in general. Researching the BISP program, we found that AI is not a catch-all solution. Although we are addressing a key part of the problem people face when accessing support, there is still a lot of work to be done to help people on a larger scale. Roshni is aimed at a very well-defined target audience. If it helped even one mother understand a message she would otherwise have ignored, if it stopped one family from falling for a scam, if it kept one child's stipend from slipping away, it would be worth it. But AI tools assume a level of digital literacy that, realistically, many people in rural areas may not have. To truly make a difference requires more informed policymaking and structural changes in how we think about "help" in general. AI can play an important role in those changes, but it needs accompanying, human-led action.

What's next for Roshni

We are excited about the problem we have identified and the potential of our solution to make an impact in the Pakistani community, particularly for its novelty. For widespread adoption, though, we know we will need to expand it. We plan to extend Roshni to more government support programs beyond BISP, while keeping the same focus: helping people understand the specific messages that decide whether help reaches them. We also plan to consult with specialists. In our primary research we identified some experts who have studied the BISP program, and we think their feedback could give us expert-level backing and advice to make the tool more effective.