Skip to work
fig.00the maker

ThomasPeng

Design Engineer

I design and build AI systems end to end: agent orchestrators, fan-out verification pipelines, and the products they ship. Everything here is real and running. Go ahead and explore it.

See all the work ↓Get in touch
real systems, running
quorum.thomaspeng.caopen live ↗
Quorum

Agent orchestrator: fans out finder agents, then K=3 skeptics per finding. Adversarial verification cut false positives 27.8% → 0%; 3/3 real bugs on a held-out module, 0 false positives. Every call traced.

fig.01the approach

Most people draw the thing or build the thing. I do both, end to end, fast.

A designer’s eye, now pointed at shipping software. I take an idea from a blank file to a live, polished product without a hand-off: the interface, the copy, the engineering, the deploy. The tools changed. The standard did not.

fig.02selected work

Things I designed
and shipped.

Every one is live. Click any frame to open the real product.

quorum.thomaspeng.calive ↗
Quorum — An agent orchestrator that verifies its own findings.
fig.012026

Quorum

An agent orchestrator that verifies its own findings.

A task-aware orchestrator: it routes each job to the cheapest capable model tier, fans out finder agents across files, then submits every finding to K independent skeptics before trusting it. K=3 verification cut false positives from 27.8% to zero, and it caught 3 of 3 real bugs on a held-out module with none false. Every agent call is traced into a UI that reads like a product.

  • Agent orchestration
  • Adversarial fan-out
  • Trace UI
  • Python
Open Quorum ↗
roastly.thomaspeng.calive ↗
Roastly — An honest roast and scorecard for any landing page.
fig.022026

Roastly

An honest roast and scorecard for any landing page.

Paste a URL and get a sharp, specific critique and a graded scorecard. The hard part was making it fun to read and genuinely useful at the same time. People share their scores, which was always the point.

  • Roast engine
  • Shareable scorecard
Open Roastly ↗
pageforge.thomaspeng.calive ↗
PageForge — A GitHub repo becomes a launch page.
fig.032026

PageForge

A GitHub repo becomes a launch page.

Point it at a repository. It reads the code, writes the copy, designs the layout, and deploys a finished page. There is no designer-to-developer handoff here, because there is only me.

  • Repo to launch page
  • LLM pipeline
Open PageForge ↗
fig.03also shipped

More from the workshop.

04Skill-Tuning CouncilA 4-proxy council that votes before any self-improvement ships.Agent orchestration
05NameVaultBrandable startup names, generated and checked.Naming
06SigilAn npx tool that writes Claude Code commands for your repo.Developer tool
07AI News BriefDaily AI news, curated by an agent, with an Atom feed.Agent
08MCPHubA directory of MCP servers for the agent ecosystem.Directory
09Roux AgencyA web studio that ships a live demo before the first call.Studio
fig.04method

How I build.

  1. i

    Idea to live in days.

    Not quarters, not committees. A working, deployed product is the first artifact, not the last.

  2. ii

    Built in Claude Code, with agents.

    Autonomous pipelines, multi-agent fan-out, and self-tuning skills do the heavy lifting. I direct; the system executes.

  3. iii

    Design and code from one head.

    No hand-off, no spec drift. The person who chooses the typeface also writes the deploy script.

  4. iv

    Shipped to infrastructure I run.

    Every product here runs on a server I own and maintain, each on its own domain. Not a free tier that sleeps when nobody is looking.