What Is Doza Assist?
Doza Assist is a local-first AI editing assistant built specifically for documentary editors. The origin story matters here: the developer has 15 years of cutting documentaries and kept running into the same problem — cloud transcription services require uploading footage that legally can't leave the edit suite. So he built the alternative. Everything runs on the user's machine. Nothing goes to the cloud.
Currently available as a manual setup (Python, Ollama, model downloads). A signed Mac app with one-click install ships in approximately 90 days. MIT licensed and fully open source.
Editorial DNA Engine
This is the feature that separates Doza Assist from generic AI transcription tools. The Editorial DNA engine analyses an editor's finished projects and learns their individual cut patterns — pacing, selection logic, structural preferences. When applied to new footage, it produces selects shaped by that editor's actual work rather than generic LLM output.
Multiple profiles per editor are supported — useful for editors who cut differently across genres or clients. This is a genuinely novel approach to AI-assisted editing that no cloud tool currently offers.
FCPXML Round-Trip
The workflow is elegant. Export FCPXML from Final Cut Pro, drop it into Doza Assist, the app reads synced audio from disk via the parsed XML, AI runs the selects pass, round-trips back as FCPXML. A 4KB XML file replaces what would otherwise be an 80GB ProRes export. Zero media movement. Deep ingestion with multicam and sync-clip support means complex documentary timelines with multiple cameras and plural audio sources are fully supported.
Export targets Final Cut Pro, Adobe Premiere, and DaVinci Resolve. OpenTimelineIO is on the roadmap as a cleaner cross-NLE abstraction layer.
Dual-Engine Transcription
Primary transcription runs on NVIDIA Parakeet — faster on Apple Silicon. Whisper is the fallback for edge cases Parakeet handles less well. The dual-engine routing is automatic, not manual. For editors with large interview libraries, the speed difference on Apple Silicon hardware is meaningful.
Local AI Analysis
AI selects run via Ollama serving Gemma 4 (Apache 2.0 licensed — the developer explicitly swapped from earlier Gemma versions for licensing cleanliness). For editors who prefer a hosted model, an optional Claude API backend is available. All inference stays local by default — no API calls, no usage fees beyond the user's own hardware.
Client Review
Client review sharing runs via Cloudflare Tunnel, self-hosted from the user's machine. No middleman storage, no third-party platform. For editors working with sensitive documentary subjects — journalists, NGOs, legal teams — this is not a minor feature. It's the architecture that makes the tool legally usable in those contexts.
Pricing
| Tier |
Cost |
Notes |
| Open Source |
Free |
MIT licensed. Manual setup required (Python, Ollama, model downloads). |
| Signed Mac App |
TBA (~90 days) |
One-click install, Apple Silicon. Pricing at launch TBA. |
Who It Is For
Documentary editors working with footage that legally cannot be uploaded to cloud services — journalism, legal, NGO, sensitive subject matter. Independent editors who want AI-assisted selects without cloud subscription fees. FCP-native workflows where FCPXML round-trip eliminates the need to export media. Privacy-first post-production teams who need client review without third-party storage.
Limitations
Apple Silicon only: Currently Mac-only. No Windows or Linux support at this stage.
Manual setup: Until the signed app ships (~90 days), setup requires Python, Ollama, and model downloads — significant friction for non-technical editors.
Solo builder: Built and maintained by one developer. Post-launch support bandwidth is an unknown.
Pre-release: The one-click Mac app is not yet available. Current state is developer-friendly, not editor-friendly.
Verdict
Doza Assist is solving a real problem with a sharp, domain-specific answer. The Editorial DNA engine is genuinely novel — no other tool learns individual editor cut patterns and applies them to new footage. The FCPXML round-trip with zero media movement is the right architecture for legal and privacy-sensitive documentary work. The setup friction is real until the signed Mac app ships, but for technically capable editors who need local-first AI selects today, this is the only tool in the category. One to watch closely at launch.