-
-
AGAR
Inspiration
Hi, I’m Abhay! I’m a student at Leland High School and I’ve been working with AI for almost six years – from robotics and computer vision to frontier research at Stanford and UCLA. I believe AI models like ChatGPT and Gemini can make us smarter by helping us discover knowledge, perspectives, and ideas we can’t find alone. However, the opposite occurs in practice: studies show AI use correlates with lower critical thinking, test scores, and creativity. The stakes are high: AI can either help our country’s next generations become more capable and creative, or stop them from ever learning how to think. To avoid the latter, we must define what responsible AI usage looks like for those growing up with these tools. Plus, AI can genuinely be transformative in the classroom, with the right guardrails.
So, to facilitate responsible student AI interaction, I built Agar.
What it does
Agar helps students who need personalized support and teachers who face burnout by creating interactive, personalized AI study partners. Powered by Gemini 3.0 Flash, a backend AI agent identifies teacher-specified acceptable methods using an uploaded assignment and reference materials. Then, a student-facing AI model helps students scaffold thinking and understand misconceptions, while keeping them aligned with their teacher’s intended outcomes and never just giving answers away.
I chose Gemini because it's fast but accurate and maintains performance across long conversations. Using American models ensures student data privacy and aligns with national security priorities regarding foreign AI.
How we built it
Almost immediately after starting, scheduling conflicts forced my teammate to drop out, leaving me with a barely-started codebase and the full burden of development, pilots, and iteration. I realized the only path forward was focusing on what mattered most – and that meant developing alongside real teachers and students.
First, I partnered with a history teacher to test Agar in his classroom. As he helped me gather feedback and refine app features, I expanded testing to include 14 teachers across 2 San Francisco Bay Area High Schools. Throughout January, I deployed new versions each morning, monitored database logs as teachers and students used Agar across subjects, grade levels and schools, then fixed what broke or made students struggle. When interviewed, teachers using Agar consistently reported their students could better explain core concepts in assessments.
More technically speaking, the app is a NextJS React application built with help from Google Antigravity+Gemini CLI and hosted on Vercel.
Challenges we ran into
The largest technical challenge I faced was AI response consistency. During early testing, Agar would randomly return improperly formatted citations or give students false information, even when instructed to do otherwise. To fix this, I created a dataset of 100+ student ‘jailbreak’ interactions and tested new prompting strategies. I found that XML-based prompts (which are closer to how AI models are instructed during their training) dropped answer fold rates from 75% to 8% and entirely eliminated citation formatting errors.
Accomplishments that we're proud of
I'm really proud of how receptive teachers have been of the app. I initially created Agar to satisfy an intellectual curiosity, but almost every teacher I've tested the app with has wanted to continue using Agar in their classroom. I've put a lot of time into making every part of the website – even microscopic things like the order in which items move around when a new message is sent – and I'm really glad that it's effects are being seen and real classrooms are being impacted in positive ways by something I've made.
This is a little smaller, but another thing I'm really proud of is that I finally figured out how to properly host a website. For a long time I've been afraid of hosting real websites because of charges associated with hosting, domains, databases, etc, but ConvexDB makes the database processes really easy (especially when paired with Google Cloud's Cloud Run for GPU accelerated tasks like OCR or sentiment analysis) and Vercel makes hosting very easy.
What we learned
The biggest thing I've learned is that promoting promotes responsible/ethical use of AI is the responsibility of developers, not users. There's a lot of stigma around AI (especially in education) because many worry that AI can just be used to get answers, which is a completely valid concern. But the solution shouldn't be telling AI users to 'be careful'; rather those who develop AI products must ensure that what they build will default towards encouraging ethical and responsible behaviors. I attempt this with Agar, and as I continue conducting AI research and building AI applications, I'll stay true to this philosophy.
What's next for Agar: Responsible, Curriculum-Aligned AI for Classrooms
Looking forward, I hope to expand Agar’s service. I’m currently working with San Jose Unified School District to approve Agar for district-wide usage, and connecting with interested educators across the country. To support wider student needs, I also hope to integrate more features like real-time speech models (Gemini 3.0 Realtime Native!) and image-generation models (Nano Banana/NB Pro!).
Log in or sign up for Devpost to join the conversation.