AI-powered 3D voxel knowledge map for children with learning disabilities

Problem  •  Solution  •  Research  •  Architecture  •  Getting Started  •  Docs

Built by Armaan Priyadarshan, Samarjit Deshmukh & Kavneer Majhail
Freshmen at Dartmouth College

Meno is named after Plato's Meno dialogue, in which Socrates demonstrates that a young, uneducated slave boy already possesses latent mathematical knowledge. He just needs the right questions to draw it out. Socrates guides the boy through a geometry proof not by lecturing but by asking, and the boy arrives at understanding on his own.

That's the thesis of this project: every child already has the capacity for understanding. A learning disability doesn't mean the knowledge is absent. It means the usual path to it is blocked. Meno's job is to find where the blockage is, make it visible, and help the adults in a child's life ask the right questions.

Learning disabilities affect roughly 1 in 5 children, yet most go unidentified until years of struggle have compounded into anxiety, academic failure, and eroded self-confidence. Dyscalculia alone affects 5 to 7% of school-age children (roughly the same prevalence as dyslexia) and is dramatically underdiagnosed. Beyond math, children may struggle with working memory, attention, spatial reasoning, or language processing in ways that are invisible to the adults around them.

Current screening tools are paper-based, require trained administrators, and produce a binary "at risk / not at risk" label with no actionable detail about where the breakdown is or what the child's strengths are. Parents and teachers are left with vague report cards and no concrete path forward.

We wanted to create something that could catch learning difficulties earlier and show the adults in a child's life exactly what is happening, not just that something is wrong.

Meno is a two-platform system (one for the child, one for the parent) that builds a living picture of how a child learns. It quietly screens for learning difficulties through themed games, maps the results onto a 3D voxel brain sculpture, and surfaces both strengths and areas that need support so every adult in the child's life can see what's really going on.

The child plays through themed game sessions with four task types drawn from cognitive neuroscience:

Every response silently records accuracy, response time, hesitation, and answer changes. The child never sees clinical language, just themed emoji and encouragement.

The parent sees a voxelized brain sculpture inside a translucent head silhouette. Each brain region maps to a knowledge-graph category (Number Sense, Number Relations, Arithmetic, Working Memory, Spatial Reasoning, Language & Number). Voxel color encodes mastery on a continuous gradient:

🟢 Green = mastered    🟡 Amber = developing    🔴 Red = needs support    ⬜ Gray = not yet assessed

The Behavioral Risk Profile runs a calibrated logistic regression classifier and produces:

• A risk score (0 to 100%) with per-feature contributions showing what's driving it
• An LLM-generated narrative (GPT-4o) translating the model output into plain-language findings and recommendations
• A confidence check (Mahalanobis out-of-distribution detection) that downgrades confidence when the child's profile is unusual relative to the training data
• A downloadable screening report (PDF) a parent can bring to a learning specialist

The goal is to show not just where a child struggles, but also where they're strong, so the conversation with a teacher or specialist starts from a complete picture rather than a deficit label.

The behavioral features Meno captures are drawn from peer-reviewed research on learning disabilities and number processing:

The initial demo classifier is a logistic regression trained on a published 64-subject dyscalculia cohort with leave-one-out cross-validation:

The architecture is designed to extend beyond dyscalculia. The screening layer supports four categories (dyscalculia, attention, working memory, general learning difficulty) and the knowledge graph can grow to cover any domain the AI decomposes. Model weights are checked into the repo as JSON. No Python runs at inference time.

Child plays games --> child_responses table --> Feature Aggregator (TS)
                                                        |
                                                        v
                                               Trained Classifier (JSON weights)
                                                        |
                                                        v
                                               GPT-4o Narrative Layer
                                                        |
                                                        v
                                        GET /api/parent-summary/[studentId]
                                                        |
                                                        v
                              Parent dashboard: voxel brain + knowledge graph + screening

The two platforms are physically segmented via Next.js route groups (app/(parent)/, app/(child)/). They share Supabase tables but no React providers. The child never sees clinical language; the parent never sees gamification.

pnpm install
pnpm dev

Open http://localhost:3000 to pick a platform.

# Profile with learning difficulty signals (risk ~91%)
curl -X POST -H "Content-Type: application/json" \
  -d '{"persona":"dd"}' http://localhost:3000/api/seed-demo

# Typically-developing profile (risk ~11%)
curl -X POST -H "Content-Type: application/json" \
  -d '{"persona":"td"}' http://localhost:3000/api/seed-demo

Then visit /parent to see the dashboard update.

uv run --project scripts/train_classifier scripts/train_classifier/train.py

Vision: Help every child get what they need for their circumstances by making their learning profile visible, concrete, and actionable for the people around them.

Demo scope: The initial demo targets dyscalculia (K-3 number sense) because that's where our personal experience and training data are strongest. The architecture is general: the screening layer already supports four categories (dyscalculia, attention, working memory, general learning difficulty), the knowledge graph can grow to any domain the AI decomposes, and the classifier is designed to be retrained on new cohorts as data becomes available.

Goal: Give parents and teachers a complete picture of how a child learns, not just a deficit label. The sculpture shows strengths and gaps together. The screening report is something concrete a parent can bring to a learning specialist to start a real conversation.

Non-goal: Diagnosis. Meno is pattern detection. Every screening surface carries the disclaimer. The risk score comes from 64 training subjects and has not been validated for clinical use.

Built with care at HackDartmouth XI
Meno - because every child deserves to be understood, not just tested.