Personal AI Agent with Voice

I built a personal assistant around a single agent loop: the model is offered a set of tools, decides which to call and in what order, and keeps going until it has an answer. Every capability is one function plus its description, so the assistant grows by registering tools rather than by adding branches.

It started with web search, then the ability to open a page and pull out what matters, which is the difference between reading snippets and reading sources. Persistent memory came next through mem0, so it recognises me across sessions instead of restarting cold every time.

The job search is where those pieces compound. It parses an uploaded resume, searches for live openings, scores each one against that specific background, and emails the shortlist. Because memory was already wired in, the resume is remembered too, which quietly makes the next capability possible.

The last layer removed the screen. A Twilio number is attached to an ElevenLabs voice agent that calls the same endpoints the web app uses, so I can phone the assistant, have it greet me from memory, ask it to run the job search, and hang up while it works in the background and emails the results. Chat and phone are not two products with two histories. They are one assistant reachable two ways, sharing one memory, which is the point at which it stopped feeling like a demo.