Cognitalk

architectual diagram
Brand logo
User Interaction

Inspiration

AI is a powerful tool for self-reflection, but would you upload your most private journal entries to a third-party server? We didn't think so. Cognitalk was inspired by the belief that you shouldn't have to trade your privacy for the power of AI. We wanted to build a truly personal assistant that is on your side, and only your side, by ensuring your sensitive data never leaves your personal computer.

What it does

Cognitalk is a 100% private, command-line AI journal assistant. It allows a user to:

Securely Add Entries: Write and save journal entries to a local text file that is never uploaded.
Ask Questions: Chat with your own history in a natural, conversational way to recall memories or find specific information.
Analyze Themes: Go beyond simple Q&A. The analyze feature uses gpt-oss to act as an AI psychoanalyst, identifying recurring themes, patterns, and emotions in your writing to provide deep, personal insights for self-reflection.

How we built it

Cognitalk is built on a secure, privacy-first client-server architecture:

The Brain (Server): A Python Flask server runs in a private Google Colab environment. It uses Ollama to serve OpenAI's gpt-oss-20b model. This backend handles all the AI reasoning.
The Bridge (Tunnel): We use ngrok to create a secure, temporary tunnel to the Flask server, allowing the client to communicate without exposing the server directly.
The Interface (Client): A local Python script runs on the user's PC. It manages the journal file and the user interface, which is enhanced with Colorama for a polished and user-friendly experience.

This design guarantees that the AI model only ever sees the text of a single prompt and never has direct access to the user's files.

Challenges we ran into

Our biggest challenge was ensuring system stability and data integrity. We solved two key problems:

Unicode Errors: Early versions crashed when encountering special characters or emojis in the journal. We solved this by rigorously implementing UTF-8 encoding in all file I/O operations.
Ollama Initialization: The AI model would sometimes fail to start before the server was ready. We re-architected the server script into a robust, all-in-one process that ensures Ollama is downloaded, running, and has the model pulled before the public-facing server starts listening for requests.

Accomplishments that we're proud of

We are incredibly proud of building a fully functional, end-to-end AI application that is both useful and safe. Our key accomplishments are:

The Privacy-First Architecture: Successfully creating a system where the user's private data is physically separated from the AI model.
Agentic Reasoning: The analyze feature is a huge win. It showcases gpt-oss's ability to perform complex synthesis and reasoning, not just basic Q&A.
A Polished User Experience: Even as a command-line tool, the app is intuitive, provides clear feedback, and uses color to be visually engaging.

What we learned

This project was a deep dive into the power of local, open-weight language models. We learned:

The Viability of Local AI: gpt-oss combined with Ollama is an incredibly powerful stack for building real-world applications that don't need to rely on cloud APIs.
The Art of Prompt Engineering: The difference between our simple Q&A prompt and the more complex "psychoanalyst" prompt for the analyze feature showed us how a well-crafted prompt can completely change the model's capabilities.
The Importance of Secure Design: Building the client-server architecture taught us how to design systems that are powerful but still respect and protect the user.

What's next for Cognitalk

This is just the beginning. Our vision for Cognitalk is to expand it into a full, private OS assistant.

Short-Term: Package the client into a simple, downloadable application for Windows and macOS.
Long-Term: Extend the core architecture to handle more than just a journal. We envision a voice-controlled agent that can analyze local documents, browse the web, and automate tasks with pyautogui, all while maintaining the same 100% privacy guarantee. Cognitalk is the foundation for a future where your AI agent works for you, and only you.