Creator @OpenObscure — On-device privacy for AI agents. Focused on distributed systems, representation engineering, and local LLM privacy.
On-device privacy firewall for AI agents — intercepts LLM prompts and responses in real time, encrypts PII with FF1 Format-Preserving Encryption, and scans every LLM response for manipulation via a cognitive firewall. Zero cloud dependency. Designed for Enchanted, RikkaHub, and any OpenAI-compatible AI agent.
| Layer | Component | Purpose |
|---|---|---|
| Detection | TinyBERT-4L-312D INT8 + CRF | Named entity recognition, low-latency on-device |
| Encryption | FF1 Format-Preserving Encryption | Reversible token substitution, structure-safe |
| Cognitive Firewall | TinyBERT cascade · 250-phrase dictionary | Scans every LLM response for manipulation — EU AI Act Article 5 |
| Vision | BlazeFace + SCRFD | Face solid-fill redaction before LLM upload |
| OCR | PaddleOCR via ONNX Runtime | Text extraction from images, PII interception |
| Safety | 5-class ViT NSFW classifier | Content classification pre-transmission |
Deployment tiers: Lite · Standard · Full — within a 275 MB RAM / 70 MB storage budget.
Integrations: Enchanted (iOS) · RikkaHub (Android) · OpenClaw (Gateway Agent)
Compliance: EAR/TSU notified · EU AI Act Article 5 reviewed · Export Control Notice included
These repositories represent earlier applied ML and computer vision research. They are no longer actively maintained but remain publicly available as reference implementations.
| Repository | Domain | Stack | Notes |
|---|---|---|---|
| computer_vision | Computer Vision | C/C++, Python, OpenPose, DLIB, Keras, TensorFlow | Object detection, tracking, face recognition, gesture & emotion recognition |
| DNN_Hyperparameter_Optimization | AutoML / Distributed ML | Python, Distributed TensorFlow | Neural network hyperparameter optimization via random search over distributed PS/Worker architecture |
| kalman_filter_multi_object_tracking | Computer Vision | Python, OpenCV | Multi-object tracking using Kalman filter estimation |
| image_compression | Numerical Methods | Python, Jupyter | Image compression via SVD, QR, and LU decomposition |
Note: These projects were built against TensorFlow 1.x / Keras 2.x era tooling. APIs and dependencies may be outdated.
- Current: Creator, OpenObscure · Director, Data Security & AI Privacy — CNA Insurance
- Focus areas: AI privacy engineering · LLM security · On-Device ML · Cognitive Firewall Research · Privacy-Preserving Systems
- Earlier work: Embedded systems, Distributed ML infrastructure, Computer Vision
