ai agents that observe your screen, learn your habits, and proactively assist with digital tasks. not a chatbot — a system that understands context across applications and acts on your behalf. the beginning of what eventually became this whole setup.
the original vision was simpler: just capture screenshots and let the ai figure out what you're doing. then it started suggesting things. "you've been in this spreadsheet for 40 minutes, want me to summarize the data?" then it started doing things. "you have a meeting in 10 minutes, i've prepped the relevant docs." now i have an ai that manages my server, deploys my code, tracks my health data, and argues with me about creatine supplementation.
lifelog.ai was the first domino. the idea that your desktop context — what apps you have open, what you're looking at, what you were doing 10 minutes ago — could feed into an ai that actually helps instead of just responds. every other ai product at the time was a chat window. you type, it answers. lifelog was the opposite: it watches, it learns, it acts without being asked.
the tech stack was python + screen capture + vision model + llm. capture screenshots every 30 seconds, feed them to a vision model to extract context ("user is in a spreadsheet", "user is reading an email from john"), store the context timeline, and use an llm to suggest actions based on patterns. the hard part wasn't any individual piece — it was making the suggestions feel timely and relevant instead of random and annoying. timing is everything. a suggestion at the right moment feels like magic. the same suggestion 5 minutes later feels like noise.
the project evolved from there. lifelog became the template for everything that followed — the esp32 presence tracking, the server automation, the health data integration. all of it follows the same pattern: observe, learn, act. the ai isn't a tool you talk to. it's an environment that knows you.