Case study · 01
Bolo
A Mac app that reads anything on your screen out loud — with the entire neural voice running on your device.
Where it started
Bolo began as the simple version: select text anywhere on macOS, hit a hotkey, and a cloud model reads it to you. Groq made that path fast and pleasant. The itch was the other path — the on-device voices you could run locally sounded robotic next to the cloud ones, and I wanted local to be a real choice, not a compromise.
The port
The voice I wanted locally was Chatterbox-Turbo — a three-stage neural pipeline (a T3 transformer, S3Gen, and a vocoder) that existed in Python on Apple’s MLX framework. I ported all three stages to native MLX-Swift. No Python runtime on the user’s machine, no bridges — the model graph rebuilt in Swift, stage by stage.
Proving it correct
A model port fails quietly: nothing crashes, the voice is just subtly wrong. So the port is validated by six numerical parity gates that compare the Swift pipeline’s outputs against the Python reference at each stage. Drift shows up as a failing number in CI, not as a listener’s vague sense that something is off.
Making it fast
4-bit quantization and a KV cache, tuned for Apple Silicon, keep generation realtime on a laptop. The result is an on-device voice that holds its own against the cloud path sitting next to it in the same menu.
What shipped
A public, MIT-licensed menu-bar app: system-wide hotkeys, Accessibility-API capture of whatever you’ve selected, and a choice between the Groq cloud voice and the fully local one. The interesting part isn’t the app — it’s that the local option exists at all.
Currently open to full-time roles. email me — or go talk to my AI.