I research trust in AI. Then I build the system and live inside it.

Clarence is a self-hosted multi-agent memory system on a single SQLite database: structured memories, an entity-fact knowledge graph, semantic vector search, scheduled jobs, and public telemetry. Governance is single-writer by design: one agent writes durable memory, and every other agent submits structured handoffs that a review filter accepts or rejects.

The researcher's move: before reranking shipped, a 40-query retrieval benchmark ran against the live database. Hit@1 went from 19 of 40 to 34 of 40, and reranking still shipped as an opt-in sidecar, not the default. Evaluation discipline, applied to my own infrastructure.

The case study also names what is not solved. Visibility of system status, Nielsen's first heuristic, remains the main open failure, documented in public rather than hidden.

“A portfolio case study that only shows what worked is a sales document.”

The organism at the top of this page runs the same math this instrument plays: a carrier, a modulator, and a room full of signals deciding how they meet. SensorSynthFM is an iPad FM synthesizer that treats motion, touch, and room energy as modulation sources, my Kent State MS UX capstone, built in Swift with AudioKit and SwiftUI.

The build history is public and evidence-gated. The microphone went from “seems active” (JUL 1) to a causal, stable two-minute route (JUL 15) only after dedicated diagnostics. A 36-cell modulation matrix, nine sources and four targets, is signed, installed, and launched on a physical iPad. When the route slider collided with the iPadOS bottom gesture zone, the fix stayed on an unmerged branch until it could be validated.

Clarence writes code fast. I own the musical judgment, the UX direction, and every public claim. AI can write code quickly. It cannot decide whether a mapping feels musical.

“Code existence is not the same as a validated interaction.”

I co-founded the Pittsburgh Ableton User Group in 2015 and led it for a decade, serving as direct liaison to Ableton's international team. I can distinguish between a design decision and a design failure because I understand the ecosystem from the inside.

The suite applies five methods across that ecosystem: a stakeholder analysis of Live (2024), Nielsen heuristic evaluations of Note (2025) and Learning Music (2025–26, severity-scaled), a cognitive walkthrough of Learning Synths (2026), and a COGA-grounded cognitive accessibility evaluation of ableton.com and Live 12 (2026). Five methods, one converging pattern.

It lands on fixes, not complaints: icon labels with a Show-labels toggle, as Push already does, and a First Session simplified mode for the person opening Live for the first time.

“Ableton designs for insiders, and that works until someone shows up for the first time.”

The anchor artifact is a VPAT 2.5 conformance report on NPR.org addressing all 56 WCAG 2.1 A/AA criteria using axe DevTools, WAVE, HeadingsMap, and custom Python static-analysis scripts. Three criteria are documented Does Not Support: 1.4.1 Use of Color, 1.4.3 Contrast (Minimum), and 4.1.2 Name, Role, Value. The sharpest pair is 74 issues on the NPR home page against 7 on a story page. Same organization, different discipline.

The frontier is cognitive accessibility. The COGA-grounded Ableton evaluation covers ADHD, dyslexia, working memory, and executive function, the barriers nobody files a ticket about. Next: open-sourcing an audit-to-VPAT reporting tool seeded from the NPR scripts.

This page practices what this section preaches: visible focus everywhere, reduced motion honored, motion pausable by button, and no interaction required to reach portfolio information.