HiveMind AI

• 

  example of the one-layer model (twolayer will be shown)

The initial project idea was to create an app based on generative AI that acted like a ‘copilot’ so to speak, one that would be able to access a user’s screen to take screenshots that would be analysed instantly based on contents.

We wanted to run our AI model locally so it would be suitable to help small-medium-sized businesses to make decisions with sensitive data without needing to share it with a third party. With a rise of usage of AI chatbots, it’s important to ensure that these features remain accessible to all, regardless of whether they can afford proper data protection. We ran our model locally on a W6800 GPU.

To make our app stand out from others, we then considered the creation of different ‘profiles’ that a user could switch between, which provided the AI additional context in a given situation to aid in decision-making. We wanted this app to have a very intuitive layout, so our initial design made use of a lot of tick box options as well as a segment to attach PDFs.

We knew that in order for our app to aid in decision making, we needed it to interpret the same prompt from different perspectives. This is how the agent concept came about, with a singular profile being able to host several agents that each give a different perspective on the decision being made depending on their focus. To keep the response concise, we decided that the app should also be able to create an evaluative synthesis of all the different agents, leading the super agent to be born. This, essentially, is a suite of specialised ‘agents’ who analyse all the different aspects of a decision and are then joined by another ‘super’ agent. We even went on to experiment with two layers of agent summarisation. This approach also allows us to summarise over 100k words within a minute using one local gpu - which is ridiculous.

Built With

• deepseek
• electron
• llama
• node.js
• ollama
• prompt
• python
• react
• vector