Chris Cardoza

Shoutout to NVIDIA for Parakeet. Doza Assist runs dual-engine transcription locally on the editor's Mac, and Parakeet is the primary engine. It's fast, accurate, and runs beautifully on Apple Silicon. When you drop a 90-minute documentary interview into Doza Assist and get back a word-level timestamped transcript in minutes without anything leaving your machine, that's Parakeet doing the heavy lifting. NVIDIA built a transcription model that competes with cloud APIs while running entirely on local hardware. That's what makes local-first AI editing possible, not theoretical.

Doza Assist needed a way to run local LLMs on a Mac without asking editors to configure model weights, manage Python environments, or touch a config file. Ollama solved that completely. It's the layer that lets a documentary editor run Gemma on their own laptop the same way they'd open any other app. The "nothing leaves the laptop" promise only works because Ollama made local inference simple enough that non-engineers can actually use it. Massive piece of infrastructure for anyone building local-first AI.

Powered by Gemma running locally through Ollama. Doza Assist needed an AI model that could run story analysis entirely on an editor's Mac with no cloud calls, no per-token costs, and no licensing complications for a commercial app. Gemma 4's Apache 2.0 license made it the clear choice. Documentary editors work with sensitive footage. Nothing leaves the laptop.

I'm a documentary editor, not an engineer. Claude Code is the reason a solo filmmaker could build a production-grade Mac app with local AI transcription, multi-NLE export, and a full Flask backend without writing a single line from scratch. It's not autocomplete. It's a co-engineer.