AI safety researcher and builder at Mac Labs AI.
I work on interpretability, unlearning, abstention and epistemics, world models (JEPA-style), and how models behave toward users. A longer list of projects is on my portfolio.
Research participant, SPAR Fall 2026 (moral shadowbanning in conversational AI).
Researching JEPA-style world models for time-series event forecasting in a research collaboration, including how event-probability forecasts should be calibrated when event rates change, and building on KITE. In progress.
Unlearning vs. Hiding (Lexsi AI Safety and Alignment Hackathon, Track 2b). Audits whether RMU unlearning erases or just hides facts in Tiny Aya (3.35B). Site · Paper · Code · Video
A Linear Answerability Signal That the Model Does Not Act On (participant, Apart AI x Epistemics sprint). A linear probe finds a signal for answerability in Phi-3.5-mini's hidden states that the model's output does not follow, with probe AUROC 0.906 at layer 18 on SQuAD 2.0. Site · Paper · Code · Video
DocsNavigator MCP. An MCP server with a Gradio app that answers questions over documentation files, using full-text search, a language model, and OCR for scanned PDFs and images. Project page · Space · Code
Docs Navigator MCP: SUSE Edition (SUSE Hack Week 25). Searches, summarizes, and answers questions over SUSE, Rancher, K3s, and RKE2 documentation with local open-source models. An independent project, not an official SUSE product. Project page · Space · Code
HealixPath (Hugging Face MCP 1st Birthday, with DevRubenArevalo). An MCP health assistant that runs on free open models; visitors sign in with Hugging Face. Project page · Space · Frozen submission · Code
17 public hackathon projects since August 2025.
More than 1,000 commits in 2026 across AI engineering, data engineering, research, and developer tooling.



