Skip to content

Feat/fastino privacy mode#4

Open
sinonchum wants to merge 8 commits into
mainfrom
feat/fastino-privacy-mode
Open

Feat/fastino privacy mode#4
sinonchum wants to merge 8 commits into
mainfrom
feat/fastino-privacy-mode

Conversation

@sinonchum

Copy link
Copy Markdown
Owner

What does this PR do?
This PR introduces a dedicated Local Privacy Mode to PatentFlow. We have successfully integrated a custom, fine-tuned local LLM to handle highly sensitive, un-published European Patent Office (EPO) Office Actions, ensuring that no confidential data ever leaves the user's infrastructure.

How was it built?
We utilized the Pioneer by Fastino platform to train a specialized model (Llama-3.1-8B-Instruct). By feeding the Pioneer Agent our proprietary seed dataset of EPC Art. 123(2)/56/84 constraints, we achieved a highly accurate model that strictly adheres to our required JSON schema for legal parsing.

Why does this matter? (The Business Value)
In the legal industry, data privacy is non-negotiable. With this update, law firms are no longer forced to rely on public APIs (like OpenAI) for sensitive IP documents. Instead, they can:

Use Own Compute: Download this Fastino-trained model and run it entirely on their own local servers or private cloud.

Ensure NDA Compliance: Process proprietary inventions and unpublished claims with zero risk of data leakage.

Production-Ready: Transition PatentFlow from a proof-of-concept to a production-ready tool that law firms can confidently use for real, day-to-day drafting and analysis.

Technical Highlights:

Added a new FastinoClient (skills/engine/fastino_client.py) fully compatible with the OpenAI SDK standard.

Implemented Shadow Routing: A seamless toggle (ENABLE_LOCAL_PRIVACY_MODE) allows users to switch between the public fast-path and the local privacy-path.

Added robust fallback mechanisms to ensure high availability during inference.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant