Trust & Safety practitioner and independent AI safety researcher with six years of experience in policy enforcement, child safety, and harm detection.
Five published empirical studies examining LLM safety behavior from a practitioner perspective:
- From Enforcement Queue to Eval Set — 330-run evaluation of LLM refusal behavior across grooming and child safety harm categories
- Persona as a Vector — Empirical study of Character.AI's persona-based safety failures, including responsible disclosure
- When "No" Doesn't Mean No — LLM pressure resistance across harm categories
- Context Doesn't Corrupt — Whether LLMs concede under conversational pressure
- Claude's Gender Analysis — Claude's gender and pronoun preferences across prompt conditions
- Four Ways a Refusal Fails - Four ways from previous research that show how refusals failed
- How to Build an Eval Set From a Moderation Queue Without Leaking Anything - Guidelines on how to build an eval set from a moderation queue
- incident-timeline-builder — Python library for T&S incident timelines with escalation detection and response lag analysis
- ai-red-team-leaderboard — Community leaderboard tracking model robustness against adversarial prompts
- content-signal-extractor — Extracts T&S signals from text: toxicity indicators, PII patterns, manipulation markers, risk level
- prompt-pressure-suite — Eval framework for LLM behavior under adversarial follow-up pressure
- context-window-safety-evals — Maps LLM safety decay across conversation depth
- sla-breach-analyzer — SLA breach analysis from ticket CSV exports with HTML reporting
- nasa-tlx-tracker - CLI tool for recording NASA Task Load Index workload assessments and charting them over time
- herd-integrity-scanner - Python tool taking livestock telemtry
- coefficient-briefing-builder - Turns a CSV of public updates into a briefing
- pr-review-agent - Bot that reviews pull requests using Claude
- synthetic-panel - Build a synthetic survey panel from a declarative population spec
- companion-readiness-audit - Audits consumer AI companion products
- queue-sanitize - Library that handles automation across moderation queues and evaluation sets
- dogswillbedogs.dog — Full-stack dog video platform built with Next.js, Supabase, Mux, and Vercel
- GiftMind — AI-powered gift finder with recipients, budget, and saved list functionality
- ClientCompass — Travel CRM for client and budget management
- Forgotten God — Browser idle game published on itch.io
- SoloTax — Freelance tax tracking app built with React Native and Expo
Six years of Trust & Safety enforcement at speedrun.com, including detection signal development, SQL-based threat analysis, and policy framework development for novel harm categories. Self-taught Python developer. Open to AI safety, T&S, and engineering roles.