NLP
A theoretical reconstruction of the Claude Mythos architecture, built from first principles using the available research literature.
Reverse proxy for Claude Code that anonymizes sensitive pentest data (IPs, hashes, credentials, hostnames, PII) before it reaches Anthropic. Dual-layer detection: local Ollama LLM + regex safety net, with per-engagement vault and self-improving feedback loop.
Learn LLM internals step by step - from tokenization to attention to inference optimization.
Karpathy-style LLM knowledge base for Obsidian. Clone, run Claude Code, start building your second brain.
Fine-tune Gemma 4 and 3n with audio, images and text on Apple Silicon, using PyTorch and Metal Performance Shaders.
A Claude Code skill that designs and builds high-converting questionnaire-style app onboarding flows — modelled on proven conversion patterns from top subscription apps like Mob, Headspace and Noom
Inference repo for Falcon-Perception and Falcon-OCR model, early-fusion, natively multimodal, dense Autoregressive Transformer models.
TurboQuant: Near-optimal KV cache quantization for LLM inference (3-bit keys, 2-bit values) with Triton kernels + vLLM integration
Local meeting transcription → Obsidian vault. No cloud, no API keys.
Lightweight, provider-agnostic Python LLM library — one API for OpenAI, Gemini, Anthropic, Groq, Mistral, Cohere, Azure, Bedrock & Ollama. Vision, streaming, tools, batch.
Simulate anything, for $1 & less than 10 min — Universal Swarm Intelligence Engine
Transcribe microphone and computer audio to markdown.
Personal project on Rust aimed to help understand foreign language better. Uses VAD+Whisper to transcribe, then translate according to the custom dictionary.
80,433-trial study of context-window sycophancy across 6 LLMs (4B–72B). Behavioral ratchet effect, correction injection mitigation, phase transition analysis. Code, data, and preprint included.
Fully local meeting transcription with speaker diarization, AI summaries, and PDF output