Back to selected work
Flagship product / Founder & Engineer

Full Circle — An AI Team That Shows Its Work

An AI team for small online businesses. Agents handle the inbox, content and follow-ups, and every job comes back with a record of what was asked, what ran, who checked it and what it cost.

Visit fullcircle.software ↗
AI AgentsSaaSNext.jsTypeScriptPostgresHuman Approval
The Full Circle workspace: a list of AI agents on the left, and an Inbox Manager reply waiting for approval with its sources, review score and cost.
The agent workspace, as shown on fullcircle.software. A drafted reply waits for approval; the panel on the right lists its sources, the independent review and the cost. The business shown is a fictional example.

The problem: An AI agent that says it finished is not the same as work that is finished. A small business owner who hands their inbox or their content to an agent has no way to tell a good draft from a confident mistake, and no record of what the agent looked at or what it cost.

The product: Full Circle gives a small online business a team of AI agents for the inbox, content and follow-ups. The difference is what comes back with each job: proof. It is not open yet. The site is taking a waitlist, with no launch date published.

One job, around the circle

Every job follows the same six steps, and each step leaves a record:

  • Asked. A job opens with the task and what counts as done.
  • Ran. An agent does the work. Each model call is traced.
  • Checked. A reviewer from a different AI lab reads the result. If it fails review, the job says so.
  • Sources. Every lookup is on record: which email, which calendar, which page.
  • Delivered. You approve it before anything goes out.
  • Cost. The exact cost, itemised by call, the reviewer's included.

The decisions that matter

  • Nothing sends by itself. Agents can read what they need right away. Anything that sends or changes something waits for your approval, with Approve, Edit and Decline on the draft.
  • A second lab checks the first. The reviewing model comes from a different AI lab than the model that did the work, so one model's blind spot is less likely to pass as a pass.
  • Unknown stays unknown. If a model call has no trace, or a delivery was not confirmed, the job shows it as not confirmed instead of claiming it.
  • Running to the end is not passing. A job that failed review, or never produced a draft, says so in plain words.

What I built

The whole product: the marketing site, the agent workspace, the job record and proof panel, approvals, billing in credits, and the connections to the apps a business already uses. It is a Next.js and TypeScript application on Postgres.

Where it stands

Full Circle is waitlist only. The workspace shown here is the example on the public site, with a fictional business. There are no customer results to report yet, and none are claimed.

The Full Circle home page: The AI team that shows its work, with a waitlist sign-up.
The public site. Waitlist only; no launch date is published.