Design lead who ships. I build the system, the team, and the thing itself.

Fourteen years across design systems, delivery and strategy, from startups to the Fortune 500. I build the thing, and I build the team that builds the thing — lately with AI agents doing real work in the loop, inside harnesses I design: the context they get, the memory they keep, the tools they use, and the review gate before anything ships. Right now that's Scout, a tool that turns raw input into structured data.

Selected work

Designed the AI harness architecture behind Sparq's delivery

Role: Architect and spec author. Built the Staffing harness.

Problem: one-off prompts leave agents without the context, limits and review they need.

What I did

  • 4 harnesses as one graph: 58 nodes, 68 connections, 8 cross-harness handoffs
  • One anatomy per harness: context, tools, human review gate, 3 memory layers
  • Feedback loop at every decision gate: outcomes rewrite the rules
  • Interactive map: click any node to trace its inputs and flows

What changed

  • New harness starts from a template, not a blank prompt
  • Human review gate on every harness before client, ship or merge
  • Handoffs are explicit inputs, not convention

Built an AI staffing harness that sizes consulting engagements

Role: Led the capability model. Designed and built the harness.

Problem: teams were staffed by gut feel and corrected mid-engagement, the costliest time to be wrong.

What I did

  • 26-capability model as the agent's semantic memory, separate from job titles
  • Agent drafts a seat range with a stated reason per line; a person confirms
  • Code does the scope math and segregation-of-duties checks; people keep the judgment
  • Predicted vs. actual staffing logged per engagement and fed back into the rules

What changed

  • Mid-engagement corrections: target 50% fewer
  • Proposal turnaround: target 2 days, down from 1 week
  • Every seat traces to a stated reason
  • Seat list reused by Proving Ground and Build

Designed and built VERA, an AI agent that drafts deliverables with sources attached

Role: Sole designer and builder.

Problem: stakeholders couldn't tell evidence from confident guesses, so the loudest claim won.

What I did

  • VERA agent plus Stacks tool: evidence attached to every claim
  • Evidence ledger with a validated-or-fiction review gate
  • Unsupported claims flagged, not deleted
  • Prototyped in code from day one; the prototype became the product

What changed

  • Deliverable assembly: days to hours, across 3 engagements
  • 0 unsourced claims shipped

Built a cross-platform design system for Swift, React Native and web

Role: Design lead. 2 engineers. Built tokens and first 40 components myself.

Problem: three platforms, no shared standard, and every design decision remade and drifting.

What I did

  • DTCG tokens through Style Dictionary into Swift, React Native and CSS
  • Operating model: semver releases, same-day Figma-to-JSON, explicit platform handoff
  • One release cut 233 speculative tokens to about 60

What changed

  • Design-to-build rework down about 40%
  • 4th product UI shipped in 6 weeks vs. a 16-week baseline
  • Adoption still above 90% after a year
  • Rebrand touches 1 file, not 40

Proof

Case studies are testimony — you have to take my word for it. This is evidence instead: the actual calls I made, the actual repo, and how I think in writing.

A DECISION LOG

  • Price capabilities at a level, not titles.

    Why

    Titles hide the work. On an enterprise client engagement, an AI & ML title built systems and UI, and a Product Strategy title ran delivery. Pricing by title would have staffed the wrong people.

    What it cost

    Harder to sell than a title-first estimate, because clients think in roles. It also needs self-assessments that don't exist yet.

  • Agents never get a seat, and their savings stay unproven.

    Why

    Counting agent help without the review it creates understaffs exactly the engagements that lean on agents hardest.

    What it cost

    The model reads high against what agentic teams deliver. Sales will want the savings in the price; it waits for timesheets.

  • Ship the iPad app gated, not blocked.

    Why

    A designer can explore design and AR on day one. Pricing features stay inactive until an admin saves one category, item and variant, so onboarding friction drops to near zero.

    What it cost

    It's a two-state product, and nobody has defined what "inactive" looks like on the iPad: greyed, hidden or placeholder numbers. That gap lands on the designer mid-session.

  • AI proposes, the rules engine disposes.

    Why

    AI emits only schema-validated proposals and never touches geometry or pricing. That guarantees buildability and makes every interaction auditable.

    What it cost

    No generative range at all. Schema failures reject silently, so a brittle schema becomes dead air the rep can't diagnose in the field.

  • Dropped sequential staging for parallel definition.

    Why

    Meetings kept getting cancelled or cut short and async homework sat for weeks. Waiting on one validated layer before the next would have stalled the project.

    What it cost

    No built-in safety of validating one layer before opening the next. Several foundational decisions stayed provisional at once.

  • Built the Hub instead of documenting it.

    Why

    A coded prototype on the client's own stack showed how each role really uses it, and surfaced gaps and gray-area ownership earlier than documents would have.

    What it cost

    We built ahead of settled structure. The retro named too little scaffolding upfront and limited early buy-in on technical constraints.

  • Let AI hold the relationships between unsettled layers.

    Why

    AI in Miro linked roles, tasks, intent clusters and navigation, so the team could work downstream while foundations were still being validated.

    What it cost

    It failed where decisions were ambiguous and reinterpreted instead of extending. The fix: lock decisions once made, and let AI extend, not reinterpret.

This site, built in the open

No case study for this one — just the repo. Every prototype, every rewrite, every rejected direction is a real commit, with an AI co-author in the log. Read the history instead of my summary of it.

View the commits ↗

From the notes

AI, context, and the shifting meaning of intent

In a human-built prototype, every element carries intent: we can point to a margin, a color, a label, and say “that's there because…” With AI-generated prototypes, that chain breaks — the artifacts look intentional, but aren't necessarily meaningful. The job shifts from owning every pixel to interpreting what AI makes, and turning the noise into direction.

Read the full article on LinkedIn ↗

Notes on building

The opinions behind the work. Open one to see where it comes from.

How I lead

Through the work. I grew a design practice from four to nine, promoted two designers to senior, and never stopped shipping myself — tokens, prototypes, the occasional pull request. The team's output is the portfolio; mine is making that output better and faster than it would be without me.