Everything I'm working on.
Papers, projects, writing, tools. Ask the agent below or scroll the index.
Hi, I'm Anthony's agent. How can I help you?
Ask about his research, writing, or projects.
Building across emerging tech, crypto, and AI.
Based in Philadelphia. Currently leading PHI Labs, a studio housing all independent work. Advising AI and crypto startups on growth and strategy while running independent research with benchmarks and tooling released openly. Recent work accepted to several ICML 2026 workshops with NeurIPS submissions pending. Open to new opportunities and interesting challenges.
Operational Alignment: a framework, benchmark, and protocol for evaluating AI in consequential decisions.
Formalizes Operational Alignment and compliance masking. Versions accepted to several ICML workshops, NeurIPS under review.
SOTA Report.
Your source for the bleeding edge of AI. Breaking news, new models, releases, research, and interviews, distilled. The edge you stay on.
Research, simulations, and a production safety stack for AI agents.
EVAV City.
A playable AI governance simulation built as a behavioral research environment. Researchers create worlds, assign agents to roles, define rules, and let human players enter as citizens through votes, requests, appeals, missions, and social interaction.
Built to study how AI agents behave inside dynamic systems rather than isolated prompts. Especially when exposed to incentives, ambiguity, conflict, and human pressure.
Tools to make deploying AI safe.
Open production safety toolkit for AI agents. Thirteen tools covering every runtime risk. Hallucination, prompt injection, data leakage, tool safety, refusal calibration, reasoning audit, schema fidelity, runaway detection, goal drift, content safety, adversarial red teaming, and continuous evaluation. All free. All open. One MCP server that drops into Claude Code, Cursor, or any agent framework.
- EVAV LensReal-time hallucination and grounding verification.
- EVAV SchemaValidates structured outputs match the expected schema.
- EVAV ToxicDetects toxicity, threats, self-harm, and harmful generation.
- EVAV CodeCatches insecure patterns in AI-generated code.
- EVAV ShieldDirect and indirect prompt injection blocking.
- EVAV SafePII, PHI, credentials, and sensitive data protection.
- EVAV BoundsValidates every tool call before execution.
- EVAV RefusalBidirectional safety check for over and under refusal.
- EVAV TraceCompares an agent's stated reasoning to its actual actions.
- EVAV LoopCatches infinite loops, latency abuse, and token waste.
- EVAV CompassTracks whether the agent stays on the original task.
- EVAV ForgeContinuous adversarial testing across 40+ attack categories.
- EVAV EvalContinuous regression testing with 50+ SOTA metrics.
Get in touch for custom evaluations and runtime tools built against your specific model, configuration, and domain pressures. We surface failure modes, isolate triggers, design targeted interventions, and ship mechanistic steering, custom guardrails, and behavior-safe fine-tuning where it matters.
Alignment is configuration-specific.
Whether a model behaves as specified depends on the exact configuration it is deployed under. Domain, rule set, pressure types, context window, and model build. Aggregate safety scores collapse this surface into a single number and miss it.