The Infinity AI BuildFest: Building Roots AI and Reaching the Finals

On Friday, June 12, 2026, our team stepped into the high-energy environment of The Infinity AI BuildFest at Brac University, Dhaka, Bangladesh — a hackathon built around one big question: how can AI make learning genuinely better?

We came in with a problem, a stack of ideas, and a tight clock. We left as finalists with a working product called Roots AI, an AI-powered learning platform for EdTech. This is the story of how that day unfolded.

The challenge: AI for EdTech

The theme was open-ended but focused — build an innovative AI solution for education. The judging panel was looking for products that didn't just demo well, but actually addressed real pain points students and teachers face every day.

We zeroed in on a problem every student has felt: most learning tools tell you what to study, but none of them tell you why you're stuck. A learner hits a wall on a topic, but the platform has no idea whether the gap is in the current lesson or three prerequisites back. We wanted to fix that.

That observation became the seed of Roots AI.

What we built: Roots AI

Roots AI is an AI-powered learning platform that diagnoses the root cause of a learner's confusion instead of just serving more content. It rests on three core ideas:

• Concept MRI — a diagnostic engine for knowledge. The AI scans a learner's interactions and identifies the specific knowledge gaps holding them back, then maps the unmet prerequisite concepts that need to be filled in first. It's the difference between "you got question 7 wrong" and "you're stuck because you never actually internalized how exponents work, and that's blocking everything downstream."
• Multimodal learning inputs. Students don't only learn from text. Roots AI accepts an image of a textbook page or problem and converts it into a structured, solvable problem inside the app — no retyping, no lost context.
• Voice-based Feynman Technique assessments. To verify that a student has actually mastered a concept (and not just clicked through it), the learner explains the topic out loud. The AI evaluates the explanation and confirms whether the mental model is real, or whether gaps are still hiding underneath.

The goal wasn't to replace teachers. It was to give students a study companion that can see the gap a teacher would see — and surface it before frustration turns into giving up.

How we built it in a day

Hackathons force brutal prioritization, and that's a good thing. We made a few decisions early and stuck with them:

1. A thin vertical slice beats a wide, broken demo. We picked one full user journey — learner lands, gets a Concept MRI diagnosis, fills a prerequisite gap, proves mastery with a voice explanation — and made it actually work end-to-end.
2. LLMs in the loop, not in the way. We used Gemini (orchestrated through langchain) as the reasoning layer and PgVector for the concept-graph retrieval that powers the diagnostic engine. The UX kept the human clearly in control: the AI diagnosed and suggested, the student decided what to learn next.
3. Ship the demo, not the architecture. We reached for tools we already trusted — Next.js 16 and TypeScript for the app, Chakra UI for the interface, Firebase for auth and persistence — so we could spend our hours on the diagnostic experience, not on plumbing.

The best part of a hackathon isn't the code you write — it's the code you decide not to write.

The moment of recognition

Being named a finalist at The Infinity AI BuildFest was a genuinely proud moment. The pool of teams was strong, the judges asked sharp questions, and the bar for "interesting" was high. We got to walk through Roots AI's reasoning, demo the personalization in real time, and hear feedback from people who actually build in this space.

It was a clear reminder that good ideas, executed with care and shipped on time, can stand shoulder-to-shoulder with much bigger teams.

Lessons I'm taking with me

A few things this hackathon reinforced for me:

• Start with a real frustration. The strongest pitch I gave all day was the one that started with "here's what actually went wrong when I tried to learn X."
• Constraints are a feature. A 24-ish-hour clock, a single demo flow, and a small team forced cleaner choices than I'd have made with a week.
• Diagnosis before content. In EdTech specifically, "AI-powered" is no longer impressive on its own — the win is in understanding the learner before you generate anything for them. Concept MRI is the part of Roots AI I'm proudest of.
• Demo the feeling, not just the feature. Watching a learner's voice explanation get a genuine "yes, you've got it" from the AI landed harder than any technical deep-dive.

What's next for Roots AI

The hackathon is over, but the project isn't. The pieces we built — the Concept MRI diagnostic engine, the image-to-problem multimodal pipeline, and the voice-based Feynman assessment loop — are all worth turning into a real, usable product. We'll be tightening the diagnostic accuracy, running it past actual students, and figuring out which slice of EdTech we want to go after first.

If you want to follow along, the easiest way is to keep an eye on my LinkedIn — that's where I post build logs and updates.

Wrap-up

The Infinity AI BuildFest was a great reminder of why hackathons are worth doing: tight timelines, kind strangers, strong coffee, and a single weekend that produces a working product you'd otherwise keep deferring for months.

Massive respect to the organizers, the judges, and every team that showed up and built. See you at the next one.