I map failure patterns across complex systems and build the security boundaries required to stabilize them.
I'm the founder of Project Navi - an open-source AI security company focused on zero trust architecture and mathematical governance. Before that, I spent seven years in behavioral health research, coordinating peer-support programs across 500+ organizations and publishing on workforce collapse in the APA Psychiatric Rehabilitation Journal.
The through-line is failure analysis. The same structural breakdowns I studied in human systems - drift under pressure, coherence loss at scale, collapse when governance is bolted on instead of built in - show up in AI deployments. So I started building infrastructure to prevent them.
Machine-verified proof that self-sustaining coherence configurations exist in nonlinear elliptic PDE systems - formalizing, for the first time, the mathematical conditions under which autopoietic closure can emerge. Lean 4 against Mathlib v4.28.0. Fifteen theorems proved, zero sorry (no unfinished proofs), 3000+ build jobs clean. Where Mathlib lacks infrastructure (Schauder estimates, maximum principles, sub/super-solution theory), the axiom boundary is explicit and auditable via a single typeclass. Algebraic core proofs were synthesized using Aristotle (Harmonic) - human architecture, machine search, compiler as final arbiter. The underlying theory is in the paper.
Machine-verified proof that the (u,v)-flower graph has hub distance u^g and log-ratio convergence to log(u+v)/log(u) - the quantity that equals box-counting fractal dimension in the physics literature (Rozenfeld, Havlin & ben-Avraham 2007). Lean 4 against Mathlib v4.28.0. Zero sorry, zero custom axioms. Explicit SimpleGraph construction with graph-theoretic distance proof. This is the ground truth that navi-fractal calibrates against.
Deterministic input sanitization for untrusted text in LLM pipelines. Strips homoglyphs, invisible Unicode, null bytes, template injection, and path traversal vectors. Zero dependencies. Python 3.12+. Live on PyPI.
AI-powered PR review agent with a deterministic security rule engine. Indexes your codebase into a knowledge graph for context-aware analysis. Runs security rules before the model touches the diff. Structured findings, severity scores, pass/fail verdicts. Works with any model.
Spec-driven repo scaffolding that ships CI, security scanning, code review, and release pipelines in a single command. Jinja2 engine with 7 template packs. Define your standards once, apply them everywhere. Python 3.12+. Live on PyPI.
Audit-grade fractal dimension estimator for complex networks with refusal semantics - if the evidence for power-law scaling isn't there, you get a machine-readable refusal, not a meaningless number. Sandbox method with four quality gates: R² threshold, AICc model selection, curvature guard, slope stability. Zero dependencies. Calibrated against networks with exact analytical dimensions from fd-formalization. Python 3.12+.
- OpenHands: Disclosed a CVSS 9.1 security vulnerability; wrote and merged the fix into
main. Contributed defense-in-depth SecurityAnalyzer ensemble. - Mathlib: Contributed SimpleGraph.ball (open metric ball).
- NIST and NCCoE: Submitted responses on AI agent identity, authorization, and adversarial prompt detection (Zenodo).
The tools are the entry point. Underneath them is a mathematical governance framework built on three foundations:
Differential Symbolic Calculus (DSC) - Measurement framework for detecting coherence-exploration coupling in dynamical systems. Instrumentation, not simulation.
World Model Capital (WMC) - Trust-allocation system adapting Recovery Capital frameworks from behavioral health to AI governance. Trust capital for machine cognition.
Relational Emergent Ontology (REO) - Philosophical foundation where intelligence is event, identity is rhythm, ethics is resonance.
You don't need to understand any of this to use the tools. But if you're curious - I started here because the principles that help people recover help systems stay whole.
- Security: security@projectnavi.ai
- Licensing: legal@projectnavi.ai
- PGP:
/.well-known/pgp-key.txt(fingerprint:402E C296 1A72 CBFF 63B8 FEE9 A42A 76A1 C696 FF08) - Sponsor my work: GitHub Sponsors
Machine cognition, human values.
The knowledge is free, the community is open. If you wish to support our mission, buy a t-shirt. 🐘