Most AI never leaves the demo. I build the part that runs in production.

The check a vibe-coded project skips on its way out the door.

Problem

Anyone can ship a working product in a weekend now. Almost nobody audits it first. Generated code goes live carrying exposed keys, missing security headers, no indexing rules and pages that fall over on a real connection. The tools that catch that properly are priced for enterprise, and the free ones hand back a score with no reasoning, so nobody knows what to fix first.

What I built

A multi-tenant SaaS that runs the full audit as a background job and returns written findings rather than a number. Crawling and analysis sit behind a queue so a slow site never blocks the app, and Claude turns raw results into prioritised, plain-language reasoning you can act on before launch, not after the first incident.

• 3audit domains: security, SEO, performance
• 5pricing tiers, self-serve checkout
• Asyncqueue-backed crawl pipeline

A whole agency runs on one system instead of four tools and a spreadsheet.

Problem

Agencies run people, clients and payroll in separate tools that never talk. Answering a question as basic as what a client actually costs to serve means a manual reconciliation nobody has time for, so nobody does it and the agency prices work on instinct.

What I built

Team, client and payroll management on a single data model, so hours, cost and delivery roll up on their own. One record per person and per client, no re-keying between systems, and the margin on a client is a number the system already knows.

• 4 → 1tools consolidated
• 3modules, one data model
• 0manual reconciliation steps

The one I'm not talking about yet.

Problem

Every team now has AI that can answer. Almost none have AI that can finish. A model finishes a sentence and hands the work straight back, so the messy middle of a job, the part with the tool access and the judgement calls and the twelve steps nobody wrote down, still lands on a person. That gap is where the real cost sits, and it is not a prompting problem.

What I'm building

An agent platform aimed squarely at that gap. Agents that hold context across a long task, act on real systems rather than describing what they would do, and are accountable for an outcome instead of a reply. It is the largest thing on this page and the one everything else has been practice for. Details when it ships.

• Agenticbuilt to finish work, not answer questions
• Long-horizoncontext held across an entire task
• Quietmore at launch