IntelliStream
AI-Powered Cognitive Profiling for Stream Selection
Stealth-gamified cognitive profiling platform that predicts academic stream fit for Indian Class 10–11 students through behavioral gameplay analytics — replacing the single high-stakes aptitude test with continuous, naturalistic measurement.
Indian students choose a life-defining academic stream at 14–15 based on a single pen-and-paper aptitude test — a one-time, self-report snapshot that can't capture how cognition develops over time or map onto modern career pathways beyond Science, Commerce, and Humanities.
A stealth-gamified platform where students play purpose-built cognitive mini-games across 8–12 sessions; behavioral signals — response latency, decision revisions, error recovery, cross-session consistency — are aggregated into a cognitive feature vector and classified by XGBoost into ranked stream probabilities.
- Six-phase cognitive taxonomy (~31 dimensions) spanning Speed, Memory, Reasoning, Verbal, Numerical & Executive Function
- Stealth-gamified portfolio of 62+ mini-games with anonymised display names, removing evaluative framing
- Dual-track behavioral signal capture — global event-listener RT proxy plus a millisecond-precise stimulus/response API
- XGBoost prediction engine (~8ms inference, 6.2MB model) with 1-hour result caching and ranked stream probabilities
Measures how a student thinks across dozens of naturalistic play sessions instead of what they report in one sitting. Behavioral process metrics proved substantially stronger predictors than raw game scores, validating the core stealth-assessment thesis.
Co-authored the research paper and led the technical build — full-stack platform, behavioral analytics pipeline, and XGBoost prediction engine — with a 4-member team from Galgotias University.
Built as a three-layer system: a React 18 + TypeScript game assessment layer (Vercel CDN), a Supabase/PostgreSQL behavioral analytics layer that aggregates dual-track reaction-time signals into a cognitive feature vector, and a Python Flask + XGBoost inference server (Render, Gunicorn). Trained on a cognitively-grounded synthetic dataset with 5-fold cross-validation, then validated in a school-based pilot across Grades 8–12, with stable predictions emerging after 8–12 gameplay sessions of 20–35 minutes each.
89.7% top-1 and 96.1% top-2 accuracy (macro F1 0.88) on evaluation data, with a live deployed platform (62+ games) and a school pilot across Grades 8–12 showing stable predictions after 8–12 sessions.