2026Live
KVKK Document Assistant
- Python
- FastAPI
- PyTorch
- ONNX Runtime
- Hugging Face Transformers
- RAG
- Groq
- Gemini
- Docker
- Render
Overview
A web app that finds Turkish personal data (names, ID numbers, IBANs, phone numbers…), replaces it with pseudonyms and lets you ask an AI model about the document without sending that data. Names are found by a Turkish BERT model I trained.
01 · Problem
Companies want to use large language models, but they cannot send documents with personal data to a model abroad, and ready-made masking tools have no Turkish support.
02 · Solution
Rules that validate numbers with their check digits, plus a Turkish NER model I trained on Colab for names. Questions are masked with the same pseudonyms, answers cite their sources through RAG, and every request sent to the model is shown in an audit log. The model runs on the server with ONNX, without PyTorch.
03 · Outcome
On held-out hard cases the rules found 14% of names and the model found 100%, with no value left unmasked. Questions are answered in about 0.5 seconds; the project has over 600 tests.