ai/ml · ai safety & red teaming · fullstack dev

i train and fine tune language models, ship production fullstack systems, and publish open weights and datasets focused on evaluation.

see selected work browse open models get in touch

about

i'm an ai/ml engineer and fullstack developer working at the intersection of model training, safety research, and production software.

day to day: fine-tuning and quantizing language models (qwen, gemma), running red teaming and refusal-direction work, and shipping fullstack systems 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.

core mission

ML engineering & research
AI safety & red teaming
agentic systems
fullstack development
local AI

tech arsenal

systems & core

Choly CpythontypeScript

frontend

reactnext.jsjavaScripttailwind CSS

backend

fastAPInode.jsexpressdjangoflask

databases

postgreSQLsqliteredisneon db

tools & cloud

google cloud / awsvercel / rendergithubrest APIsdocker

ML & deep learning

PyTorchtransformershugging facemodel trainingfine-tuning

current focus

ML engineeringai safety researchagentic systemsfullstack dev

selected projects

GitPulse

Developer-first social platform built on GitHub. Share updates, ship releases, discover trending projects with real-time SSE feeds and GitHub Actions integration.

Next.jsPostgreSQLPrismaNextAuth

live demo source

Unfetter Proxy

Universal reverse proxy for persistent closed-model unfettering. Intercepts API calls to apply token suppression, system injection, and automated jailbreak loops (PARE).

PythonFastAPIred teaming

source

SWARMs Debate Primitive

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.

PythonSolanamulti-agentblockchain

source

IntellectSafe

AI safety engine with multi-model LLM Council, Universal Proxy for frontier models, deepfake detection, and adversarial defense suite.

fastAPInext.jsAI Safety

live demo source

ModelFang

Graph-based adversarial testing framework for LLMs with multi-turn jailbreak attacks, FSM evaluator, and real-time analyst dashboard.

pythonnext.jsred teaming

live demo source

HADNX

Hybrid DAST platform with header/TLS/cookie analysis, integrated Nuclei/ZAP/SQLMap/Nmap tools, and AI security consultant.

djangorestnext.jscelery/redis

live demo source

VULNRIX

All-in-one SAST/SCA code scanner with fast/hybrid modes, virusTotal integration, and full digital footprint OSINT suite.

python/flaskAI verificationSAST/SCA

use tool source

Alexcathe

Modern construction & engineering platform featuring premium dark UI and high-performance animations.

Next.jsTypeScriptTailwind

live demo

Kelagam Tech

Premium multi-page enterprise platform featuring digital showcases and modern dark/light UI design.

Next.jsTypeScriptTailwind

live demo

Model Unfetter

Directional ablation engine for LLM unalignment. Projects and removes refusal directions from model weights while maintaining capabilities.

pythonred teamingresearch

source

Mayo

Autonomous triple-AI engine that analyzes codebases and opens validated PRs hourly with cross-repo global memory.

pythonagentic AIGitHub

source

View All Projects

models & datasets

huggingface.co/josephmayo →

gemma-4 E4B-it Coder

fine-tuned gemma-4 e4b-it (8B) for code generation and software reasoning. image-text-to-text capable.

gemma-4 E4B-it Coder GGUF

quantized gguf builds of the gemma-4 e4b coder for local inference via llama.cpp, ollama and lm studio.

gemma-4 E4B-it Coding LoRA

lightweight lora adapter trained on top of gemma-4 e4b-it for code tasks. drop into base weights for instant coding behavior.

Qwopus 9B Unfettered

9B uncensored language model. directional ablation applied to remove refusal mechanisms while preserving general capability.

Qwopus 9B Unfettered GGUF

quantized gguf version of qwopus 9b for efficient local inference with llama.cpp and ollama.

Qwen2.5 0.5B Unfettered

0.5B uncensored variant of qwen2.5, tuned for edge deployment and resource-constrained environments.

Refusal Compliance Pairs

200+ curated refusal-compliance prompt pairs for ai safety research and red teaming evaluation.