Six cognitive psychology principles. Eight weeks of engineering. One mission: make Indian students love learning again.
In my previous article, The Science of Joyful Learning, I outlined what 140 years of cognitive research actually tells us about how the human brain learns. The short version: almost everything about how Indian students are taught to study is wrong.
Passive re-reading beats nothing — but barely. Rote repetition without understanding creates fragile knowledge that crumbles under exam pressure. And the chronic high-stakes anxiety that surrounds JEE, NEET, and board exams actively suppresses the hippocampal activity required to form lasting memories.
Writing that article was the easy part. What I have spent the months since doing is considerably harder: building a system that actually implements those principles at scale, for real students, in Hinglish.
On April 12, we open AI Mitra to our first cohort of 200+ alpha testers. This article explains what we built, why each decision was grounded in specific cognitive science, and what we need from our early community.
The architecture of a cognitive engine
Most EdTech apps are delivery mechanisms — they get information to the student. AI Mitra is designed to be a processing mechanism — it helps the student’s brain do the work that creates durable understanding.
Every feature maps to a specific, named cognitive principle. Let me walk through them.
1. The daily check-in — affective forecasting and optimal challenge
The science: Yerkes and Dodson established in 1908 that performance peaks at moderate arousal and collapses under either too little stimulation or too much stress. Csikszentmihalyi’s flow research later showed that learning enters a deep, efficient state only when challenge slightly exceeds current skill — not by a large margin.
What we built: Every day, Mitra opens with a 60-second emotional check-in. Students rate their stress (1–4), mood, and energy. This is not journaling for its own sake — the stress signal feeds directly into Mitra’s AI context: a student who checks in at stress level 4 gets a notably different opening message than one at level 1. The AI acknowledges their state before asking a single academic question.
Crucially, this data is stored in a DPDP Act 2023-compliant audit trail. The emotional signal influences how Mitra teaches — it is not shared with anyone, including parents, without explicit consent. We built privacy-law compliance into the data model itself, not as an afterthought.
2. The SM-2 flashcard engine — Ebbinghaus’s forgetting curve, actually implemented
The science: Hermann Ebbinghaus mapped the forgetting curve in 1885. Without review, we lose roughly 50% of new information within an hour and 90% within a week. Spaced repetition — reviewing at the right moment of near-forgetting — creates far stronger retention than cramming.
What we built: When a student uploads any PDF or NCERT chapter to the Study Hub, Mitra generates flashcards and schedules them using the actual SM-2 (SuperMemo-2) algorithm — the same one that powers Anki.
Here is what “actual SM-2” means in practice: each card carries an ease factor (starting at 2.5, ranging 1.3–2.5) and a day-interval that grows or resets based on the student’s self-rated recall (0 = blackout, 5 = perfect). A card rated 5 is scheduled ~6 days later, then ~15, then ~38. A card rated 2 resets to a 1-day interval and its ease factor drops. This is not a simplified imitation — the mathematics of forgetting are coded into the model.
3. Socratic mode — the zone of proximal development in practice
The science: Vygotsky’s Zone of Proximal Development identifies the gap between what a student can do alone and what they can do with guidance. A good teacher does not solve the problem; they ask the next right question.
What we built: Mitra’s Maths and Physics tutors operate in Socratic Mode by default. The Maths prompt reads: “NEVER give the solution directly. Always ask the student to try first. Guide with the smallest possible nudge.” Forbidden phrases include “This is simple,” “Just…,” and “Obviously” — because Carol Dweck’s work on growth mindset shows these phrases actively damage a student’s belief in their ability to improve.
When Socratic Mode is active, the AI breaks every problem into its smallest sub-steps and asks exactly one guiding question per reply — never two. The student does the intellectual work; the AI holds the map.
4. Story mode — the narrative transportation effect
The science: Jerome Bruner identified two modes of thinking: logical/analytical and narrative. Most textbooks operate purely in the logical mode, but narrative transportation activates more brain regions at once, creates stronger emotional encoding, and improves recall.
What we built: When a student asks about a concept — Newton’s Laws, photosynthesis — they can switch to Story Mode, and Mitra tells the origin story of that discovery: the human moment, the confusion, the breakthrough.
There is one rule written into every Story Mode prompt: never mention JEE or NEET weightage in Story Mode. Invoking exam anxiety at the moment of wonder kills the very neurological state — curiosity-driven exploration, dopaminergic reward — that makes narrative learning work. Story Mode is protected intellectual space. The exam can wait five minutes.
5. The dual-signal weak-topic engine — diagnostic learning analytics
The science: Metacognition — knowing what you know and don’t — is one of the strongest predictors of academic success (Flavell, 1979; Hattie’s Visible Learning). But students are poor self-assessors; they confuse familiarity with understanding.
What we built: Mitra doesn’t ask students to self-assess. It detects weakness from two independent signals:
- Quiz accuracy — MCQ performance across at least 2 attempts per subject, weighted 60%.
- SM-2 recall quality — flashcard recall across at least 2 sessions, weighted 40%, with grade-awareness (an “almost” answer earns half credit).
A subject is flagged weak only when the combined signal drops below 60% and at least one signal has enough data to be meaningful. This dual gate prevents false positives — one bad quiz day doesn’t brand a student as weak in physics. The result surfaces in a Focus Areas card showing up to 3 subjects ranked by severity, each with a one-click path back into targeted tutoring.
6. The parent portal — informed, not invasive
The science: Research on parental involvement is consistent: high warmth, moderate involvement, and low surveillance is the optimal model (Pomerantz, Moorman & Litwack, 2007). Parents who monitor every score and message — however well-intentioned — create exactly the anxiety that suppresses learning.
What we built: Parents see a dashboard with their child’s check-in streak, 7-day average stress trend, total engagement points, and any safety alerts. Nothing else — no raw quiz scores, no chat transcripts, no document contents.
The safety layer deserves its own note. If Mitra’s real-time prompt-safety system detects language associated with self-harm or crisis, it does two things at once: it responds directly to the student with a helpline number, and it queues a privacy-respecting alert to linked parents. The student is never alone with a crisis moment. This has been the most careful part of the build.
What alpha testing means, specifically
We have 200+ students registered for April 12. What we need from them is not praise — we need productive failure. Specifically:
- Where does Mitra give an answer when it should have asked a question?
- Does the stress check-in feel natural, or performative after two weeks of daily use?
- Which of the 8 subject tutors (Physics, Chemistry, Biology, Maths, English, History, Geography, Economics) feels most — and least — like a real teacher?
- Does Story Mode genuinely change how students feel about a topic, or does it feel like a gimmick?
If you are a student in Class 9–12, a JEE or NEET aspirant, or a parent who has watched their child drown in exam pressure — we want you in this cohort.
Education in India does not need another platform that delivers content faster. It needs a system that respects how the brain actually works — and is honest about the gap between what cognitive science promises and what has actually been built.
We have built something real. On April 12, we find out if it works.
AI Mitra is in alpha. Features will break. Data will be treated with the seriousness it deserves. Feedback will be read personally.