Featured Projects
Ledger: catching the agent error that scored 0.93 confidence
Year: 2026 · Domain: Healthcare procurement · Capabilities: AI product design, Interaction design, Information architecture, Prototyping
A concept review layer for medicine procurement: AI agents do the work, humans review it, and autonomy is gated on what each act invented, not on confidence.
Highlights
- Hide the confidence score and show where each value came from: read, derived or assumed.
- Gate autonomy on each act with three questions: did you invent an input, who does this land on, what would it cost to undo.
- Order the queue by irreversibility, then buyer visibility, then exposure, then age.
Impact
- Agent decisions modelled: 13 in the synthetic decision log undefined 13 of 13, across 8 review shapes
- Prototype surfaces: a decision log, no interface undefined 17 live routes, no backend, 24 to 27 August 2026
- Spec tasks applied: 193 specified across 3 OpenSpec changes undefined 182 applied
Always offering a patient a way to book: IntegraMédica appointments
Year: 2019 · Domain: Healthcare · Capabilities: UX Design, Interaction Design, Design Systems, Front-end Development, Prototyping
I redesigned a clinic network's appointment booking app so every patient leaves with an appointment, in person or by video. Successful appointments rose 52% after launch.
Highlights
- Mark every day with free slots by a dot on the calendar.
- Offer a video consultation whenever no in-person slot fits.
- Keep one task on screen: a short, guided flow with a visible progress bar.
Impact
- Successful appointments: previous booking app undefined up 52%, after launch, during the pandemic
Closing the gap between design and production code with AI
Year: 2026 · Domain: AI developer tools · Capabilities: Product strategy, Interaction design, Design systems, Design engineering, AI agent orchestration
I built VortSpec 2.0 to bring the design to development gap to zero: an agentic development environment where AI designs on a real canvas and ships production code.
Highlights
- Ship a desktop app that drives the user's own local Claude Code.
- Put a real design canvas inside the app, with a .fig file as the document of record.
- Accept any design source and reduce it to one token artifact the whole pipeline shares.
Impact
- 2.0 releases: stable 0.1.40, no canvas (cut 12 Aug 2026) undefined 2.0.0-alpha.1 and alpha.2 published; 2.0.0-alpha.3 released on 20 Oct 2026, 6 to 20 Oct 2026
- Figma link rebuilt as canvas masters, one real file: one empty placeholder master per component (18 Oct 2026) undefined 305 real masters and 3,083 variants, 0 placeholders, in about 20 s, 19 Oct 2026
- Token references kept in exported code (2.0 spike): 0% with OpenPencil's own Tailwind export undefined 95.3% with the DesignIR exporter, Sep 2026 spikes
Teaching AI coding agents to build UI from specs, not prompts
Year: 2026 · Domain: AI developer tools · Capabilities: Design engineering, AI agent workflows, Design systems, Developer tooling, Technical writing
I adapted an open spec-driven template into an npm toolkit that makes AI agents write a UI spec before any code. 51 releases in six weeks.
Highlights
- Make three written specs a hard gate before any code: Component, Interaction and Page Spec.
- Ship the method as Markdown skills, installed with one npx command.
- Pilot the richest atom through all 7 steps, then fan out per tier.
Impact
- npm releases: v1.0.0 on July 1, 2026 undefined 51 versions, up to v1.26.0, by August 9, 2026
- Skills in the toolkit: 5 in the first commit undefined 20 in v1.26.0, July 1 to August 9, 2026
- Validation run on a Figma design system: Button spec set, 4 files, 824 lines undefined 39 components in one pull request, 69 files, July 2, 2026
CometVort: replacing product handoffs with one shared context
Year: 2026 · Domain: AI product development · Capabilities: Framework design, Design research, Information architecture, AI workflow design
I designed CometVort, a product lifecycle framework where people and AI agents work from one checkable shared model instead of passing documents down a line.
Highlights
- Make the object model a checkable anchor that validates code and never generates it.
- Split every gate into a machine check and a human approval.
- Scale rigour with a ladder: Patch, Feature, System.
Impact
- Published framework: planning documents in a repo undefined live site with an overview and 4 reference pages, since 24 July 2026
- Stage toolkit: no toolkit undefined 24 templates and 30 Claude Code skills across 6 stages, added 19 October 2026
From 47+ button styles to 4 for a 65,000-member union
Year: 2024 · Domain: Design System / Union Digital Platform · Capabilities: UX Research, Competitive Analysis, UI Inventory, Design Systems, Interaction Design, Accessibility (WCAG 2.2 AA), Engineering Handoff
A union's public site and member portal shared no UI standard and had never been audited for accessibility; the new system cut 47+ button styles to 4.
Highlights
- Build a full Bootstrap 5.0 design system with 28 component categories.
- Keep the existing brand and fix how it was applied.
- Show each member services based on role and job status.
Impact
- Button variants: 47+ distinct styles undefined 4 standard variants (primary, secondary, outline, ghost), Jul to Sep 2024, at design handoff
- Background colour values: 12+ ad hoc values undefined 4 token-based semantic roles, Jul to Sep 2024, at design handoff
- WCAG 2.2 AA failures: Never audited undefined 0 across 28 component categories, At design handoff, Sep 2024
One navigation model 200+ regional offices can reuse
Year: 2023 · Domain: Enterprise Website · Capabilities: UX Research, Competitive Analysis, Heuristic Evaluation, Information Architecture, Interaction Design, Visual Craft, Engineering Handoff
A US insurance broker's site got harder to navigate with every acquisition. I rebuilt its IA and components so new regions configure templates instead of commissioning redesigns.
Highlights
- Reorganise the top level around user intent: industry, service type, region.
- Keep a multi-page site with permanent URLs.
- Use one 3-column mega-menu, identical in every section.
Impact
- Critical and major usability issues resolved in the delivered design: 14 open undefined 13 resolved (93%), at design handoff, Apr 2023
- Second-level sections using a consistent navigation pattern: 34% undefined 100% in the delivered IA, at design handoff, Apr 2023
- WCAG 2.1 AA critical failures: 6 undefined 0 across all 24 components, at design handoff, Apr 2023
A portfolio that answers questions by rebuilding its own page
Year: 2026 · Domain: AI Product Design · Capabilities: AI Integration, Generative UI, Voice Interface Design, Design Systems, Full-Stack Development, Design Tooling
Static portfolios show outputs, not thinking. I designed and built a site that answers typed or spoken questions by assembling the relevant work on screen.
Highlights
- Make the canvas the main interface, not a chat panel.
- Keep a grounded homepage in front of the canvas.
- The model picks from a fixed module library; code owns layout.
Cutting time to first payrun from 4.2 to 1.8 weeks
Year: 2026 · Domain: HR / Payroll SaaS · Capabilities: UX Research, Information Architecture, Interaction Design, Usability Testing, Design Systems, Engineering Handoff
Only 34% of new clients finished self-setup without staff help; a guided task hub raised completion to 67% and cut time to first payrun by 57%.
Highlights
- Rebuild onboarding as a task hub with sequential unlocking.
- Keep a fixed task order inside each hub.
- Make the "What to have ready" screen a required first step.
Impact
- Time to first payrun: 4.2 weeks undefined 1.8 weeks, FY2024 cohort vs Q1 2026 post-launch cohort
- Self-setup completion rate: 34% undefined 67%, FY2024 cohort vs Q1 2026 post-launch cohort
- Internal touchpoints per client, first 30 days: 8.3 undefined 2.4, FY2024 cohort vs Q1 2026 post-launch cohort
Turning 502 survey answers into a profile-based discounts strategy
Year: 2022 · Domain: Fintech · Capabilities: UX Research, AI-driven Design, Design Systems
A wallet with 1 million users needed a new service that grew usage. I turned survey and analytics data into a profile-based discounts proposal.
Highlights
- Build discounts as the new service.
- Match benefits to three behavioural personas.
- Organise benefits by category with filter chips.
Cutting POS operator errors from 12% to 2.5% across 8,000 terminals
Year: 2021 · Domain: UX Design · Capabilities: UX Research, Design Systems, Wireframing, Visual Craft
Operators on a 12-option Android POS menu made errors in 12% of critical flows; a progressive disclosure redesign cut that to 2.5% in a 4-week A/B test.
Highlights
- Show 5 primary actions and move everything else behind More.
- Ship one interface for every store type.
- Use an explicit, labelled More button in the primary row.
Impact
- Operator error rate, critical flows: 12% undefined 2.5%, 4-week A/B, 1,600 redesigned vs 6,400 legacy terminals
- Settle/Close time on task: 3.8 min undefined 58 s, 4-week A/B, 1,600 redesigned vs 6,400 legacy terminals
- Settling and transaction support tickets: about 1,000/month undefined about 580/month, Phase 1 cohort, 4-week A/B
Cutting critical payment errors from 94 to 4 per million
Year: 2022 · Domain: Fintech · Capabilities: UX Research, Wireframing
Every terminal model had its own payment flow; one configurable state machine replaced them, cutting critical errors 96% and device integration from 4 to 6 weeks to 9 days.
Highlights
- One parameter-driven state machine for every device model.
- Three distinct screens for processing, timeout and declined.
- No retry on a declined payment, with a plain reason on screen.
Impact
- Critical payment errors per 1M transactions: 94 undefined 4, 60 days after launch, Jul to Aug 2022
- Median end-to-end transaction time: 45 seconds undefined 36 seconds, 60 days after launch, Jul to Aug 2022
- New device integration, design through QA: 4 to 6 weeks undefined 9 days, First new model after launch