AI Safety Research | Adversarial Discovery | Developer Tools
gabriella@kryptive.com | gabriellabaris.com
Transitioning from the CGI and VFX industry to AI safety research, leveraging creative problem-solving skills and technical expertise to contribute to responsible AI development through adversarial discovery, safety evaluation, and tool development for the AI community.
Adversarial discovery, Red teaming, Safety evaluation, Vulnerability assessment
Fine-tuning (LoRA/QLoRA), Hugging Face Transformers, PyTorch, Unsloth
Python, API Integration, Web Scraping, Data Processing, Tool Development, Pipeline Automation
• How reasoning models hide their true motivations behind fabricated policy refusals
• Uncovering policy manipulation, evaluation awareness, and infinite loops in gpt-oss; OpenAI's new open source reasoning model
• Built a professional toolkit for creating high-quality conversational datasets for LLM fine-tuning
• Implemented support for multiple export formats (ChatML, Alpaca, ShareGPT/Vicuna)
• Integrated real-time token tracking with popular tokenizers (OpenAI, HuggingFace, Mistral)
• Features auto-save functionality, progress tracking, and format-specific customizations
• Developed integration for Ollama and LM Studio
• Enabled seamless interaction with local LLMs through Discord interface
• Fine-tuned local models using Unsloth
• Created custom datasets for specialized model training
• Modeled 3D assets and environments to create animations and renderings for marketing materials.
• Character models and textures for Leia and Darth Maul for cryptoys.com digital collectibles
• Visualization artist on films including Nope, Transformers Rise of the Beasts, and Dune: Part Two
• Modeled and helped generate 3660 NFT piñatas for www.hbdnft.com