Cognito Site
Architecture

Cognito : An AI Tutor That Knows What You Should Learn Next

Inspiration

Learning has never been more accessible, yet it has never been easier to feel overwhelmed. Students consume YouTube videos, PDFs, tutorials, and online courses every day, but they are often left asking the same question after every study session:

"What should I learn next?"

We believe an AI tutor should do more than answer questions or explain concepts. It should understand a learner's progress, identify where they are struggling, and provide a practical path forward.

That belief inspired Cognito.

Cognito is an AI-first personalized learning platform built around Ajibade, an intelligent tutor designed to transform passive content consumption into active, guided learning. Instead of simply delivering information, Cognito helps learners build consistency, track progress, and receive personalized recommendations based on how they actually study.

What it does

Cognito turns topics, YouTube videos, and PDFs into interactive AI-powered classes. Learners can study with Ajibade through guided lessons, take quizzes, track their study progress, and monitor their learning journey from a single dashboard.

What makes Cognito different is its personalized learning intelligence.

Rather than treating every study session as an isolated experience, Cognito understands a learner's real progress and provides actionable guidance on what they should do next.

Using GPT-5.6, Cognito analyzes learning signals such as:

Study streaks
Total study time
Completed lessons
Weekly learning goals
Active class progress and class titles
Completion percentages and unit counts

From that context, Cognito generates personalized Study Insights, on demand, each time a learner requests them:

Progress assessments
Learning strengths and weaknesses
Prioritized recommendations
Suggested next topics to learn
Personalized weekly study plans
Daily study targets and focus areas
Motivational guidance from Ajibade

The result is an AI tutor that not only teaches concepts but also helps learners make consistent progress toward their goals.

What's new for OpenAI Build Week vs. what already existed

Cognito's core platform — authentication, class creation, Topic/YouTube/PDF ingestion, Gemini-powered lesson delivery, WebSocket lesson sessions, and progress tracking — predates this hackathon.

Built new, during the Submission Period (July 13–21, 2026), using Codex and GPT-5.6:

The full AI Study Insights feature: the authenticated /study-insights endpoint, the GPT-5.6 integration, the response contract, and the /insights dashboard page.
Backend: StudyInsightsController.java, StudyInsightsResponse.java, StudyInsightsService.java, plus supporting config changes.
Frontend: insightsService.ts, useInsights.ts, Insights.tsx.

None of these files have any commit history before this submission — that's independently verifiable with git log --follow on any of them.

We're calling this out explicitly because Cognito is a pre-existing project. Everything above is what we're actually submitting for judging; the rest of the platform is included only as context for how the new feature fits in.

How we built it

Cognito is built with Java, Spring Boot, PostgreSQL, Redis, WebSockets, and JWT authentication on the backend, with a React/TypeScript frontend.

The platform uses a multi-model AI architecture with clearly defined responsibilities:

Google Gemini powers syllabus generation, lesson delivery, and interactive teaching experiences.
GPT-5.6 powers Cognito's personalized learning intelligence layer — the Study Insights feature — helping learners understand their progress and determine what to learn next.

When a learner requests Study Insights, Cognito authenticates the request via JWT, retrieves only that learner's own study statistics and active learning data on the backend, and constructs a focused GPT-5.6 prompt from real learner context — streaks, study minutes, completed lessons, weekly goals, class titles, completion percentages, and unit counts. The response is constrained to JSON, parsed into a typed contract, and returned through a stable API. No AI credentials or raw learner data are ever exposed to the client.

Codex was the primary tool used to build this feature. Three concrete examples from the session (full detail in the repo README):

We asked Codex to implement the feature end-to-end in the existing codebase — endpoint, GPT-5.6 call, DTO, controller, and config. It shipped across 6 files in about 5 minutes, and caught an unrelated pre-existing build script bug on its own so we didn't chase a false error.
We had Codex audit its own documentation against the shipped code. It caught that our first-draft README described the JSON response with the wrong field names, and that a line claiming the feature "continuously adapts" overstated what is actually an on-demand endpoint — both fixed before this submission.
We directed Codex to fix exactly what that audit found — correcting the schema and removing stray references that didn't belong in the backend repo — in under 30 seconds.

Human-owned decisions: product scope, which learner signals to feed the model, the Gemini/GPT-5.6 responsibility split, and the security boundary keeping learner data server-side.

Challenges we ran into

Designing a Multi-Model AI Architecture

Cognito already had mature Gemini-powered teaching workflows. Replacing them simply because we were introducing GPT-5.6 would have been unnecessary and disruptive. Instead, we designed a multi-model architecture where each model focuses on what it does best: Gemini teaches, GPT-5.6 plans and personalizes.

Turning AI Responses Into Product Features

Generic AI responses are difficult to act on and difficult to present consistently in a UI. We solved this by enforcing structured JSON outputs from GPT-5.6, so every insight contains predictable sections: recommendations, priorities, weekly plans, and suggested next topics.

Protecting Learner Data

Personalized study guidance requires access to sensitive learning context. Study Insights operates entirely on the authenticated backend — learner statistics are retrieved from the JWT-backed security context, and nothing is exposed to the client beyond the final structured response.

Making Recommendations Truly Personalized

By grounding GPT-5.6 with real study behavior — streaks, progress percentages, completed lessons, weekly goals, and active classes — we generate recommendations that are practical and specific rather than generic motivational text.

Catching our own overstatements

Our first documentation pass described the feature more ambitiously than it actually behaves, and documented a response schema that had already drifted from the code. Auditing our own README against the implementation — rather than just writing it once and moving on — caught both issues before submission.

Accomplishments that we're proud of

Built a working, deployed AI tutor that transforms topics, videos, and PDFs into guided learning experiences.
Designed a multi-model AI architecture combining Gemini's teaching capabilities with GPT-5.6's personalized learning intelligence.
Shipped a personalized Study Insights engine that turns real learner progress data into actionable recommendations, end-to-end, within the Submission Period.
Built in a verification step for our own claims, catching inaccuracies before they reached judges.
Kept learner data and AI credentials server-side throughout.

What we learned

We learned that personalization is not simply about generating content — it is about helping learners make better decisions. We learned that defining clear responsibilities for AI models produces better product experiences: separating teaching from learning intelligence made the system easier to scale and improve. And we learned that structured AI outputs — and verifying documentation against real code — matter as much as the model itself.

What's next for Cognito

Save and visualize historical learning insights over time.
Track weekly study plan completion.
Turn recommendations into one-click study actions.
Introduce adaptive reminders and accountability systems.
Enable richer follow-up conversations with Ajibade based on previous sessions.
Build evolving learning profiles that improve recommendations over time.

For Judges

Live app: https://cognito.software
Test account (pre-loaded with study history so Study Insights returns real output): Email: [demo email] Password: [demo password]
Verify new work independently: bash git log --follow --date=iso-strict --format='%h %ad %s' -- src/main/java/com/cm/cognito/services/StudyInsightsService.java