The privacy filter is great, but it leaves a gap for non-technical users. Therefore, Dataiku's open source team at 575 Lab has built the Kiji Privacy Proxy. It is a complete app (MacOS/Linux/Chrome extension) that uses a fine-tuned DeBERTa model. It acts as a forward proxy, but we also have an experimental transparent proxy setup.
It detects 20+ entities, not just masks them, but also converts them back on the return trip.
The project and the entire ml stack are open source and Apache 2.0 (dataset and model on Huggingface, label review, and ml pipeline setup).
As an ML engineer at Digits’ AI team, I spent the last 2 years anxious that our verticalized approach would be obsolete once OpenAI or Anthropic released their next API. How could our specialized accounting system compete when these giants had billions in funding?
But today, our results speak volumes: we’ve built an ML system that outperforms GPT-4o by 54% in transaction categorization, achieving 93.5% accuracy through domain-specific optimization rather than raw compute. After watching our system reach 100% accuracy for some businesses through specialized learning loops, I’m convinced that small teams focusing on specific problems have a future.
The lesson for other ML engineers feeling the existential dread of competition with foundation models: domain expertise and specialization can beat generalized approaches in vertical applications. You don’t need to match their resources—you just need to solve one problem exceptionally well.
It detects 20+ entities, not just masks them, but also converts them back on the return trip.
The project and the entire ml stack are open source and Apache 2.0 (dataset and model on Huggingface, label review, and ml pipeline setup).
Repo: https://github.com/dataiku/kiji-proxy Demo: https://youtu.be/txzzY5bU2Ig