PROMPTHEAANTONIO MARTINEZ / PRODUCT CATALOGBlueDoor is a desktop preparation layer for lawyers who want to use AI with sensitive documents. Review personal information locally, replace selected details with consistent aliases and export a reviewed copy before choosing what to share with ChatGPT or another AI service.
In development for Mac with Apple Silicon. Models download separately; local workflows can then run offline. Available for private demonstrations; no public installer is offered.
Project anatomy
Legal teams need the context in contracts, case files and interviews to ask useful questions of AI. Those same materials contain client names, contact details and other sensitive information. Removing that information by hand across documents and transcripts is repetitive, and uploading an original file can disclose more than the user intended.
BlueDoor adds a local preparation and review step before online AI work. Import a document, let on-device models suggest sensitive details, then use the private map to correct detections, add omissions and choose aliases. Review and edit the document before exporting a protected PDF. Audio can also be recorded or imported for local transcription. The aim is to help the user share a more deliberate, reviewed copy while retaining useful context for later analysis. Detection can miss information and contextual clues can still identify people, so human review remains essential and anonymity is not guaranteed.
Product context
A lawyer wants help summarizing a contract in ChatGPT. In BlueDoor, they review the contract locally and replace a client name with an alias such as CLIENT_01, keeping that reference consistent throughout the document. They inspect every page, export the reviewed PDF and decide whether to attach it to ChatGPT. The original file and the private mapping are not automatically sent. BlueDoor prepares the material; the online AI service performs the subsequent analysis.
Qwen3 suggests sensitive information for the user to review. Apple Vision provides on-device OCR for scanned PDFs and images on macOS. Whisper or NVIDIA Parakeet transcribes audio, and NVIDIA Nemotron can identify speakers with compatible local models. The document editor and native export checks operate locally. After the chosen models are downloaded, these preparation workflows can run offline. Transcribing audio does not anonymize the recording or remove a speaker's voice identity.
BlueDoor is designed as a complementary preparation layer for legal workflows using ChatGPT and OpenAI. Today, its ChatGPT button only opens the browser; the user manually attaches the exported file. A proposed next step is an opt-in OpenAI API connection for summaries, document questions and drafting from a user-approved, reviewed copy. That integration is not implemented. The design goal is to keep original files and the private alias map local, with explicit review of what would leave the device. BlueDoor is developed independently.
A working macOS prototype for Apple Silicon, available for private demonstrations with public source code on GitHub. The current focus is document review, export quality and the local-to-online handoff. Public installer distribution remains pending licensing evidence, Developer ID signing, notarization and clean-Mac validation. The repository includes an internal audit describing the checks performed and their limits.
Main features
Detect sensitive details locally and replace them with editable aliases
Review the private map, correct detections and add missing information
Read scanned PDFs and images with on-device OCR on macOS
Record or import audio for local transcription and speaker identification
Edit document pages, review the result and export a protected PDF
Open ChatGPT in the browser and manually attach only the file you choose to share
Tauri 2
Svelte 5
Rust
Qwen3
Whisper
NVIDIA Parakeet
NVIDIA Nemotron
Apple Vision