ConceptCraft

Inspiration

During my Live-in-Labs fieldwork with PRADAN in Ulihatu Village, Jharkhand, I spent time with a Mundari-speaking tribal community where formal education for many students stopped at 5th standard. Around the same time, I was watching the AI education space explode with new tutors, study apps, and "personalized learning" tools — and I realized almost none of them were built for the students I'd just met. Every one of them assumed fluent English reading, confident typing, and comfort navigating a text-heavy interface.

That gap stuck with me. The people who could benefit most from AI-powered learning are often the exact people current AI tools quietly exclude. ConceptCraft started as an attempt to close that gap — not by building "yet another AI tutor," but by building one that's voice-first, grounded, and genuinely accessible from the ground up.

What it does

ConceptCraft turns any topic — a PDF, pasted notes, or a textbook photo — into a personalized, voice-narrated lesson through a five-stage learning journey:

Ground Material — uploaded content is extracted and chunked into passages, forming the source of truth for everything that follows.
Learn & Listen — a level-adaptive explanation (Beginner, Intermediate, Exam-Revision) is generated and read aloud via voice, with an English/Hindi toggle, so students who struggle to read fluently can still learn.
Adaptive Quiz — quiz questions generated strictly from the source material, with difficulty that adjusts to performance.
Teach-Back Verification — the student explains the concept back in their own words, and a semantic verification agent compares it against the real source material, praising what they got right and flagging genuine misconceptions.
Mastery & Visual — an interactive visualization for supported topics, reinforcing understanding visually.

Every single output — explanations, quiz questions, verification feedback — is grounded strictly in the student's own uploaded material. The AI is never allowed to invent facts outside the source.

How I built it

The frontend is built with React and Vite, styled with a custom warm, human-centered design system rather than a generic AI-tool template. PDF and text content is parsed and chunked client-side, then passed to Google's Gemini API for grounded generation at each stage — explanation synthesis, quiz generation, and semantic teach-back comparison.

Voice input and output run entirely on the Web Speech API, keeping the accessibility layer dependency-free and reliable without needing a paid TTS service. Every AI-generated response includes passage-level citations back to the source material, so students (and judges) can see exactly where an explanation came from.

Critically, the app also includes a local fallback engine — a rule-based generator that activates automatically if the Gemini API is unavailable or a quota limit is hit, so the app never fully breaks even without live AI access.

Challenges I ran into

This build had more real debugging than I expected for a hackathon timeline:

A live model deprecation mid-build. Partway through testing, Gemini returned a 404 for the exact model name I was using — it had been deprecated for new API keys without warning. I had to trace the actual error message, discover the replacement model alias (gemini-flash-latest), and switch to it so the app wasn't tied to a specific version that could vanish again.
A subtle chunking bug. PDF text extraction was joining all text into a single line with no paragraph breaks, which meant my RAG chunker only ever produced a single giant "chunk" instead of multiple meaningful passages — breaking grounded citation quality until I fixed the extraction logic to preserve structure.
A local fallback that couldn't generalize. My offline fallback engine was initially hardcoded around one sample topic (electrical circuits), which produced badly mismatched feedback when a student explained an entirely different subject like photosynthesis — a good reminder that "it works on my one test example" isn't the same as "it actually generalizes."
Gemini's free-tier daily quota. With only 20 requests/day on a fresh project, aggressive testing during development exhausted my quota more than once — forcing me to build (and rely on) a genuinely robust local fallback rather than treating it as an afterthought.
A teach-back dead end. Early on, if a student's teach-back attempt scored below a certain threshold, the UI offered no way to retry or continue — a real usability trap I had to catch and fix before it reached anyone else.

Accomplishments that I'm proud of

Building a genuinely grounded (RAG-style) pipeline where every AI output traces back to real source material, not open-ended generation.
The Teach-Back verification feature — a semantic comparison engine that actually evaluates whether a student understood a concept, rather than just answering their questions. This is the feature I'm most proud of: it moves the app from "explains things" to "proves you understood."
A working, dependency-light voice layer (Web Speech API) that makes the whole experience usable for students who can't rely on confident reading or typing.
A resilient fallback architecture — the app degrades gracefully instead of breaking when API limits are hit, which mattered more than I initially expected given real quota constraints during testing.
Building this entire pipeline solo, end-to-end, within the hackathon timeline — from RAG grounding to voice accessibility to semantic verification.

What I learned

I learned firsthand that "AI-powered" features are only as good as their grounding — an ungrounded or lightly-tested fallback path can silently produce confidently wrong output, and it's easy not to notice until you test with real, messy, multi-topic source material instead of a single clean example. I also learned to respect API lifecycle risk: even production model names can be deprecated with little warning, and building around aliases rather than hardcoded versions is a small design choice that pays off quickly.

Most of all, I learned that the most technically impressive feature isn't always the flashiest one — Teach-Back Verification doesn't look as visually dramatic as an interactive simulation, but it's the one thing that turns an AI study tool into something that can genuinely tell a student, and their teacher, whether they actually understood the material.

What's next for ConceptCraft

Expanding interactive visualizations beyond the current supported topic to more subjects automatically
Adding more Indian regional languages beyond Hindi for both explanation and voice narration
Building a teacher-facing dashboard so educators can see aggregated weakness patterns across a whole classroom, not just per student
Exploring offline-first support for low-connectivity rural areas, where the accessibility mission matters most
Moving to a higher-tier or self-hosted model setup to remove the free-tier daily quota constraint for real classroom use