Benchmark, linear probe and fine-tuning pipeline for visual content moderation (CLIP, SigLIP, EVA-02). B.Sc. Thesis @ Flensburg UAS.
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.
Hybrid-state streaming safety classifier and CUDA inference engine
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.
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.
Unofficial record: 284 false positives from Anthropic's safety classifier in one household's Claude Code use (Fable 5, Opus 5.5). English + Traditional Chinese.
To associate your repository with the safety-classifier topic, visit your repo's landing page and select "manage topics."