Saathi

Inspiration

The hackathon prompt asked us to find someone locked out of learning — not because of intelligence, but because of how the material reaches them. That hit close to home. Millions of students in India understand a subject perfectly well when it's explained in the language they think in, but get stuck the moment the textbook or teacher switches to English. Google Translate doesn't fix this — it just moves the same confusing sentence into a different script. It translates words, not understanding.

I wanted to build something that re-explains a concept the way a patient human tutor would — in the student's own language, not a literal translation of a textbook line.

What it does

Saathi is a lightweight AI tutor that:

1. Lets a student pick the language they understand best (Hindi, Kannada, or English)
2. Lets them choose a topic from a short, focused list
3. Takes their specific confusion as input ("Why do we flip the sign when dividing by a negative number?")
4. Sends that to an open-source LLM (via Featherless AI), instructed to explain the concept fresh, in plain words, in the chosen language — not translate a paragraph

The goal isn't to replace a teacher. It's to remove one narrow, specific barrier: the language a concept is delivered in shouldn't be what blocks a student from understanding it.

How I built it

• Frontend: Plain HTML, CSS, and JavaScript — no framework, no build step. This was intentional: the audience I'm designing for may be on lower-end devices or slower connections, and a zero-dependency page loads instantly anywhere.
• AI backend: Featherless AI's OpenAI-compatible chat completions API, calling an open-weight instruction-tuned model (meta-llama/Meta-Llama-3.1-8B-Instruct), called directly from the browser.
• Design: Kept the UI to a single, calm screen — language, topic, question, answer — so a student isn't navigating menus while confused about a math problem.

Challenges I ran into

• Scoping down, hard. My first instinct was to build something much bigger — every subject, every language, adaptive difficulty, progress tracking for parents. I had to consciously cut all of that to one feature, done properly, because a narrow tool that actually works beats an ambitious one that doesn't.
• First hackathon, first real API integration. Working with Featherless's API, understanding authentication, and debugging a 401 error taught me a lot about how API keys and account provisioning actually work behind the scenes — a genuinely new skill for me going in.
• Resisting scope creep mid-build. Partway through, I drafted a version with quiz generation and free-text topics, but it introduced a server-dependent API call that silently broke on static hosting. Catching that and rolling back to the simpler, working version was its own lesson: working and small beats broken and ambitious, especially under a deadline.

Built With

• css
• html
• javascript