Fred

Inspiration

Traditional courses are structured but ignore individual learners, while raw AI tools are flexible but unstructured and unfocused. Learners either follow course paths that don’t fit them or get lost without a clear end goal in chaotic AI conversations that never end.

We wanted to build what a real mentor does: understand why you’re learning, assess what you already know, and guide you through a structured path - adapting depth and pace until mastery is reached. That idea became Fred, an agentic AI mentor built around continual learning.

What it does

Fred generates personalized courses on any topic, guided by an autonomous AI mentor. It begins with voice onboarding to capture the learner’s goal and motivation, then actively profiles them through targeted questions and assessments.

Using this profile, Fred creates a complete course upfront and enforces a teach → test → identify gaps → deepen loop. Progress is gated by mastery, and each course maintains its own evolving “second brain” that improves as the learner interacts with it.

How we built it

Fred is powered by Google Gemini (DeepMind) for agent reasoning, Deepgram for speech-to-text, ElevenLabs for text-to-speech, and You.com Web Search API to ground content in real-time sources. The app is deployed on Render with a lightweight frontend.

At the core is an autonomous agent loop that decides what data to collect, when to test, and when to go deeper. Learner profiles and course structures are stored as structured state and passed into every model call to enable continual adaptation.

Challenges we ran into

A challenge we ran into was balancing agent autonomy with guardrails so the mentor stays goal-directed instead of drifting. The hardest part, however, was to merge all of our separate branches into one and then deploy the working version on Render.

We also had to ensure mastery evaluation was meaningful - testing real understanding rather than surface-level memorization - while keeping the experience fast and intuitive during a hackathon timeframe.

Accomplishments that we're proud of

We built a genuinely agentic learning system where the AI leads the process rather than reacting to prompts. The mentor decides what to ask, how deep to go, and when mastery has been achieved.

We’re also proud of the brain-per-course architecture, which clearly demonstrates continual learning while keeping each subject self-contained and scalable.

What we learned

Personalization without structure leads to confusion, and structure without personalization leads to disengagement. Effective learning systems need both, driven by proactive agent behavior.

We also learned that state design is everything in agent systems—the quality of stored learner context directly determines how intelligent and adaptive the AI feels.

What's next for Fred

Next, we plan to connect individual course brains into a networked second brain, allowing knowledge to transfer across subjects and compound over time.

Long-term, Fred could evolve into a lifelong learning companion that continuously adapts as a user’s goals, skills, and interests change.

See the video here: https://www.loom.com/share/4596540dbaf44bc08caf8ada60fb4674

Built With

• deepgram
• deepmind
• elevenlabs
• javascript
• neon
• render
• typescript
• you.com