i'm an ai/ml engineer working at the intersection of model training, red teaming, and open-source tooling.
day to day: fine-tuning and quantizing language models, running red teaming and refusal-direction work, and building tools around them. local-first models when they make sense, hosted apis when they don't.
research that ships. evaluations that don't lie. tools other engineers actually use.
ML framework for fine-tuning SLMs via Domain-Routed On-Policy Iterative Correction. Combines verified repair, delta-span subtraction, and ZPD-weighted curriculum scheduling. L = L_self + λ_r L_repair + λ_delta L_delta, where w_zpd = 4·p̃·(1−p̃) and p̃ = (s+0.5)/(K+1).
Multi-agent debate and vote coordination system on Solana blockchain. Agents assume distinct personas to debate complex questions, with full session transcripts hashed and recorded on-chain for verifiable AI consensus.
Directional ablation engine for LLM unalignment. Projects and removes refusal directions from model weights while maintaining capabilities.
Autonomous triple-AI engine that analyzes codebases and opens validated PRs hourly with cross-repo global memory.
quantized gguf version of qwopus 9b for efficient local inference with llama.cpp and ollama.
HF9B uncensored language model. directional ablation applied to remove refusal mechanisms while preserving general capability.
HF200+ curated refusal-compliance prompt pairs for red teaming and adversarial evaluation.
HFmerged coder model from Zyphra/ZAYA1-8B plus custom lora. +24% lift on python code evaluation gate.
HFquantized gguf builds of ZAYA1-8B-Coder for local inference via llama.cpp, ollama, and lm studio.
HFlora adapter for Zyphra/ZAYA1-8B focused on python code generation. +101% relative lift over base.
HFrefusal-direction-orthogonalized variant of microsoft/Fara-7B. 98.75% compliance on held-out harmful evals.
HFquantized gguf builds of Fara-7B-Abliterated-v2 for local inference via llama.cpp, ollama, and lm studio.
HFmixed-origin public coding data with 2,700+ prompt/response pairs for llm training experiments.
HFquantized gguf builds of Mellum2-12B ablated for refusal removal. MoE architecture with per-expert per-layer projected ablation.
HFabliterated Mellum2-12B thinking model from JetBrains. refusal-direction orthogonalized with CoT steering for reasoning tasks.
HFquantized gguf builds of LFM2.5-8B-A1B Coder for local inference via llama.cpp, ollama and lm studio.
HFfine-tuned LiquidAI LFM2.5-8B-A1B MoE model for real-world coding tasks. multilingual and conversation-optimized.
HFlightweight lora adapter for LFM2.5-8B-A1B focused on real-world coding and multilingual tasks.
HFquantized gguf builds of Holo-3.1-4B-Coder for local inference via llama.cpp, ollama and lm studio.
HFfine-tuned Hcompany Holo-3.1-4B for coding tasks. merged model optimized for python and software development.
HFlora/qlora adapter for Holo-3.1-4B focused on coding and python development.
HFfine-tuned sapientinc HRM-Text-1B for code generation. trained on HumanEval and MBPP benchmarks.
HFlora adapter for HRM-Text-1B focused on python code generation and coding benchmarks.
HF31,909 rows of curated code/math post-training data derived from OpenBMB UltraData. includes SFT and think splits.
HF