Hybrid-state streaming safety classifier and CUDA inference engine
Fix Claude Code Auto Mode classifier thinking latency, timeouts, permission denials, EADDRINUSE and port conflicts across concurrent DeepSeek/Anthropic-compatible sessions.
A design framework and evaluation set for how a model-routing safety layer should explain itself to a benign user, without disclosing the trigger to an adversary.
Purpose-built inference engine for Mistral Shieldstral 1.0 3B safety classifier. Optimized for RTX 3050 4GB VRAM. Custom CUDA kernel, INT4 quantization, multimodal vision pipeline, FastAPI moderation API.
Benchmark, linear probe and fine-tuning pipeline for visual content moderation (CLIP, SigLIP, EVA-02). B.Sc. Thesis @ Flensburg UAS.
A Claude Code skill that turns a third-party REST API into a safe, single-file Python CLI + matching runtime skill — credential never enters AI context, every mutation is classifier-gated (read / mutating / destructive / billable / financial / privilege) and audit-logged.
shiprocket-api-skill — a safe, single-file Python CLI mediating the Shiprocket API, plus a matching Claude skill. Credential never enters AI context; every mutation is classifier-gated and audit-logged.
To associate your repository with the safety-classifier topic, visit your repo's landing page and select "manage topics."