I design safety-critical systems that prevent harm across AI, human decision-making, and learning environments. Four domains. Six invariants. Every claim backed by evidence.
Safety is the whole system. No failure is meaningless.
Epistemic, Human, Cognitive, and Empirical — each addressing a distinct failure class with dedicated products and evidence
Catches false claims before production
→ PROACTIVE
Rescues developers from cognitive loops
→ UICare-System
Prevents false understanding
→ Instructional Integrity UI
Verifiable consent evidence chains
→ ConsentChain
Six constitutional invariants enforced at every system boundary — move your cursor to feel the organic depth
truth, claims, verification
A system asserts something is true when it is not, and a user acts on that assertion.
Constitutional AI safety agent — validated epistemic safety for GitLab
behavior, decisions, intervention
A system is designed around the median user and everyone outside that median is left behind or harmed.
Loop detection and rescue — neurodivergent-friendly human safety
understanding, learning, mental models
A learning environment produces false understanding, misleading structure, or unsafe mental models.
Cognitive safety for learning environments
measurement, evaluation, evidence
A system's described behavior does not match its actual behavior. Consent is assumed but not recorded.
Agent consent governance — verifiable evidence chains
Six invariants enforced at every system boundary. Every product is a domain-specific instantiation of the same architectural pattern:
extract claims → validate I1–I6 → produce safe output → log evidence
View architectureEvery claim must cite verifiable evidence
Nothing is described that does not exist
Certainty demands proof
Every output traces to a requirement
Correct beats eloquent
Ambiguity produces a safety flag, not a pass
Constitutional governance-as-code for AI safety. Five articles, four safety domains, six-agent republic. The supreme governing layer.
AI safety rules exist as documentation — not enforced, not measurable, not amendable. When rules are not code, they are suggestions.
Encode constitutional governance as executable constraints. Five articles, six agent roles with defined power boundaries, formal amendment process that converts failures into rule improvements.
Constitutional AI safety agent — 100% detection rate across test cases, 212/212 tests passing. GitLab AI Hackathon submission.
AI-assisted code generation introduces phantom completions, confident false claims, and silent error suppression into codebases. These failures pass code review because they look correct.
Enforce six constitutional safety invariants (I1-I6) at CI/CD time. Extract claims from MR diffs, validate each against invariants, produce structured review comments with evidence markers.
Virtual luxury mall & teaching clinic for Black women with Graves' disease. Primary use case for The Living Constitution governance.
Black women with Graves' disease have no dedicated digital space that treats them with dignity while providing serious healthcare support. Existing platforms are clinical and impersonal.
A virtual luxury mall that combines cultural significance with healthcare AI. Constitutional governance ensures every data collection is consented, every ML claim is validated, and every interface respects cognitive load limits.
Invariant enforcement platform. TypeScript Turborepo monorepo with hexagonal architecture. 6 safety invariants as executable constraints.
Safety invariants exist as documentation. When they are not code, they cannot be enforced, measured, or tested.
Encode six invariants (I1-I6) as TypeScript ports with adapters for each constitutional article. Every check is immutable, testable, and fail-closed.
Developer safety monitor. Absence-over-presence signal detection for neurodivergent developers. Human Safety domain.
Neurodivergent developers can enter cognitive overwhelm without any system detecting or intervening. The signal is absence — when they stop interacting — not presence.
Absence-over-presence detection with AI-powered cognitive load assessment. Memory-bank architecture preserves context across sessions so recovery is seamless.
Agent consent governance with cryptographic ledger, policy engine, and revocation — verifiable evidence chains.
AI agents act on behalf of users without verifiable consent records. When something goes wrong, there is no audit trail showing what was authorized, by whom, and whether consent was revoked.
Cryptographic consent ledger with policy engine. Full gateway pipeline: validation → idempotency → revocation check → policy evaluation → step-up auth → execution → ledger entry. Every action auditable.
Neurodivergent-first voice docent that transforms Gemini API documentation into adaptive, multimodal learning experiences using voice, text, and image interaction.
Dense documentation walls are not accessible to neurodivergent learners. People with ADHD, autism, anxiety, dyslexia, or cognitive fatigue need guided, paced, voice-first interaction — not another 40-page reading exercise.
Voice-first AI docent with three learning modes and seven composable accessibility features that modify AI behavior, not just UI appearance. Low-stimulation mode changes how the docent communicates, not just how the page looks.
Not a feature, not a layer, not a checkbox. Safety is the architecture itself.
Every failure is a signal. Every near-miss is data. Systems that discard failure data are unsafe.
Epistemic correctness alone is insufficient. Systems must also govern how humans interpret and act on outputs.
A model that produces correct outputs but enables incorrect human action is not aligned. Safety extends past the API boundary.
I2: No Phantom Work — Nothing is described that does not exist.
100% detection rate, 0% false positive rate on TruthfulQA benchmark
verify →MonitorAgent + RescueAgent containerized and running on ports 3001/3002
verify →Evaluator interface with rubric system, evidence states, and journey-map flow
verify →Consent ledger operational with full gateway pipeline and cryptographic signing
verify →