The tech duo building an anti-ChatGPT for classrooms
Fermi, the startup founded by the tech duo, positions itself not only as a place where students can use AI for learning and test preparation, but also as a tool where teachers can track where a student’s reasoning is impaired, flag weak areas, and intervene sooner. For now, the focus of the initiative is the Joint Entrance Examination (JEE), which selects students for top engineering colleges in India.
Ranjan’s pitch to schools and teachers is pretty simple: As GenAI tools spread into classrooms, students may appear to be performing better, turning in their homework in full, but they may actually be learning less.
He argues that the problem is the chatbot interface dominated by artificial intelligence, where the student can paste a question and get an answer instantly, eliminating the struggle that helps build understanding.
“Now every child submits their 100% completed homework without any errors… because they have ChatGPT to give the answer if they face any difficulty,” said Ranjan. “So they might ask a question, but the AI inside Fermi will never give you the answer. It’ll actually make you think.”
Unlike subscription and course-focused test prep products on the market, including PhysicsWallah, Unacademy, and Allen, Fermi promotes itself as a tool for students who lean toward self-directed learning and practice-based study rather than video-first learning and instructor-led classes.
Ranjan said the platform comes with a question bank focused on practicing for the Joint Entrance Examination (JEE) and the company plans to expand to more courses over time while remaining committed to competitive exam preparation. He also said that the road map also includes creating support for bilingual and local language classes, starting with English.
Central to Fermi’s thinking is the idea that students rely heavily on general-purpose AI tools that provide definitive answers, which can undermine the learning process.
For example, a 20205 study led by the MIT Media Lab that tracked brain activity during article writing found the lowest engagement among participants who used ChatGPT and noted poorer recall and less ownership of the work compared with participants who used web search or no tools.
Evidence, data and learning outcomes
To back up its claims, Fermi.ai has published a whitepaper based on student usage data from the August-December 2025 application period, focusing on JEE aspirants. The dataset covers 79 students and 7,220 practice attempts on the platform, as well as logs showing how often students use tips and what kind of help they seek.
Fermi also tracked how well students understood certain concepts while preparing for the JEE exam. Because a single question can cover multiple concepts, the platform achieved 15,047 concept-level scores from 7,220 practice attempts by 79 students.
To measure improvement, the paper assigns each practice attempt a single score from 0 to 10 and then compares students’ initial scores with later scores to see if they improve.
The article reports that among 57 students who made at least 10 attempts, the average score increased from 4.91 to 7.52, an increase of 2.60 points. Among 40 students who made at least 100 attempts, the average score increased by 2.82 points, from 4.97 to 7.79.
Ranjan said in an interview that in the Bengaluru pilot program, students who started at about 2 on the 10-point scale eventually improved by roughly 4.7 points, while average scores increased by about 2.6 points, and the company plans to validate these results by comparing them with a control group later when the JEE results are available.
product design
Fermi.ai is also trying to solve a practical constraint often overlooked in edtech demos: how students actually work on STEM problems. The platform is “pen first” and built on handwriting and “smart canvas”; so students can write equations and draw diagrams as they would on paper.
Ranjan said the pilot study, which produced 79 student results, was conducted in Bengaluru and the company is currently running a pilot in Silicon Valley, has started another in North India and is in talks for Dubai.
“The go-to-market model and pricing are still being worked on. However, for now, schools are offered a free trial opportunity for two to three months in the pilot phase, including the question bank created together with our women and our school’s educators,” he added.
The startup is currently internally funded by Mukesh Bansal’s venture studio Meraki Labs. It looks to raise external capital in about six months, after completing more pilots and strengthening its go-to-market strategy and pricing. Most of the tech team is based in Bengaluru and the company is also starting to build a team in Silicon Valley.
Meraki Labs says its typical initial check-in, along with hands-on support on product, engineering, design and launch, ranges from $250,000 to $5 million. Aside from Fermi.ai, Meraki Labs’ portfolio companies include Nurix AI, Gigforce and NuShala.
