Inspiration

• We wanted a way to turn any syllabus or course outline into a plan that actually teaches you - not just a list of topics.
• The goal of the project was to be able to upload materials, get a clear roadmap, then learn through a tutor that explains, quizzes, and moves you forward when you’re ready.

What it does

• Turns uploaded documents into a structured learning path with modules, requirements, and aligned content.
• Includes an AI tutor that teaches each requirement with a short blurb and example.
• Generates quick assessments (multiple choice and short-answer) and grades your understanding.
• Automatically advances you to the next requirement when you’ve shown mastery, with progress you can see at a glance.

How we built it

• Frontend in React + TypeScript + Ant Design for a clean, responsive UI.
• Backend in Node/Express with endpoints for the tutor, grading, and document processing.
• Two LLM roles:
  • A chat model that teaches, generates questions, and grades mastery.
  • A content model that converts raw text/PDFs into chunks, modules, and requirements.
• Guardrails like strict JSON outputs and “first-valid” parsing keep assessments stable and fast.

Challenges we ran into

• Getting consistent, valid JSON from the model for quizzes and short answers.
• Balancing speed, cost, and quality - especially for grading, which benefits from deeper reasoning.
• Handling messy inputs (short files, mixed formats) and still producing a useful roadmap.
• Making tool responses (quizzes, grading) feel natural inside the chat flow.

Accomplishments that we’re proud of

• A complete adaptive loop: teach -> assess -> grade -> auto-advance.
• Fast onboarding: you can upload content and start learning within minutes.
• A friendly UI that gives immediate feedback and easy retries.
• A modular setup that can switch models/providers as needed.

What we learned

• Clear "contract-first" prompts and tight output schemas dramatically reduce flaky behavior.
• Separating "teaching" from "grading" improves both quality and user trust.
• UX matters: inline assessments with explanations and quick retry make learning feel smooth.
• Even simple fallback logic for requirements makes first-run experiences much better.

What’s next for Cognify

• Better ingestion (diagrams, images), multi-document synthesis, and web sources.
• Spaced repetition, personalized review sessions, and deeper analytics.
• Streaming responses, hints for open-ended answers, and partial-credit scoring.
• More routing strategies and evaluation harnesses to tune quality, cost, and latency.

Built With

• antdesign
• cohere
• express.js
• gemini
• react
• supabase
• typescript
• windsurf