Working systems, on camera

A desktop app that puts a real-time AI voice persona on calls. The pipeline is live speech-to-text, a language model with persona and behaviour rules, then expressive text-to-speech, plus live vision so it can see the screen it's sharing a call with. Around it I built user accounts, a credit wallet with subscriptions, an installer, and a server proxy so customers never hold API keys.

A character that starts guarded and warms up as the conversation goes on. It has its own voice, a mood system that changes word choice without changing the voice, and safety rules that are always on. Much of the work is behavioural tuning against real call logs: cutting repetition and made-up facts, and making sure it doesn't answer a TV playing in the background.

The persona watches what's on screen during a call (a game, a video, whatever is being shared) and reacts to it out loud, in character. A vision model reads fresh screen captures every few seconds, and a separate gameplay mode tunes its comments for watching someone play. The screen capture runs through native Windows calls so grabbing frames never stalls the audio, and it only switches on with the user's consent.

Barge-in so the persona can be interrupted mid-sentence, smarter end-of-speech detection, turn-taking in group calls of 3–5+ people, speaker labels to tell voices apart, latency work on the LLM stage, and self-healing audio when devices change mid-call.

A warehouse for conversation data with a consent ledger and training-eligibility rules, automatic removal of personal names and details, curation scripts, and builders that turn calls into training-ready text and two-speaker audio datasets.

An AI teammate for Left 4 Dead 2 on a local dedicated server, with a language-model "brain", bodyguard behaviour and voice, connected to the persona. It's groundwork for AI companions that play alongside you.