Production products, shipped in weeks.
Built with Claude Code.
Not AI hype. A real engineering method: specs first, Claude Code implements, a senior engineer gates every line on auth, money, and security.
Three of our own products prove it. They are in production. Read the case studies, and open the live product.
Diagram: our delivery pipeline. A written spec, then Claude Code implements against it, then a senior engineer gates auth, money, and security, then ship.
The interesting part is what it refuses to ship.
Anyone can generate code quickly now. The part that still takes judgement is catching the thing the model got confidently wrong before it reaches production. Four from a single day, including two that were live on this site.
It defended a number I was about to "fix"
A review flagged "27 modules, 250+ tests" as stale. The pipeline proved it was correct, a day-three snapshot of a three-day build, and that the obvious correction would have published a claim that 583 tests were written in three days.
It refused to publish a contaminated eval
A 40-case safety set looked like strong evidence until three of its eleven crisis cases turned out to be verbatim few-shot examples inside the classifier’s own prompt. Contaminated by construction, so it did not ship.
It caught a bug three times larger than the search showed
A dead link read as 15 occurrences in source and emitted 45 in the build, because two components assembled the URL from a template literal. The planned check would have passed green while about 33 live pages stayed broken.
It took claims off this site
Four certification claims that were not held, and a price shipping inside structured data that no longer matched the product. Found here, on my own marketing, and removed the same day.
Most AI projects do not fail at the model. They fail at the deployment.
An MIT study of 300 enterprise AI projects found 95% produced no measurable impact on profit and loss. The models worked. What broke was the legacy database, the SSO, the data residency rule, and whether the team inheriting it could maintain it.
I have spent 4 years on the other side of that problem: embedded in a UK enterprise ticketing platform (80+ venues, billions of events a year), writing production code in a codebase I did not write, to standards I did not set.
Speed is what the method produces, not what it promises.
3,446 commits across 6 codebases in 2026, alongside the full-time role above. On 2026-08-10 I shipped to all 6 of them in one day.
| 2026 | Feb | Mar | Apr | May | Jun | Jul | Aug |
|---|---|---|---|---|---|---|---|
| CostPatrol | CostPatrol active in Feb | CostPatrol active in Mar | CostPatrol active in Apr | CostPatrol active in May | CostPatrol active in Jun | CostPatrol active in Jul | CostPatrol active in Aug |
| VowTrust | VowTrust active in Feb | VowTrust active in Mar | VowTrust active in Apr | VowTrust active in May | VowTrust active in Jun | VowTrust active in Jul | VowTrust active in Aug |
| Boody | Boody not active in Feb | Boody not active in Mar | Boody active in Apr | Boody active in May | Boody active in Jun | Boody active in Jul | Boody active in Aug |
| Trading | Trading not active in Feb | Trading not active in Mar | Trading active in Apr | Trading active in May | Trading active in Jun | Trading active in Jul | Trading active in Aug |
| brand-engine | brand-engine not active in Feb | brand-engine not active in Mar | brand-engine not active in Apr | brand-engine not active in May | brand-engine active in Jun | brand-engine active in Jul | brand-engine active in Aug |
| This site | This site not active in Feb | This site active in Mar | This site not active in Apr | This site active in May | This site active in Jun | This site active in Jul | This site active in Aug |
Fast because the method is real,
not because we cut corners.
The reason most AI-built software is unsafe is that nobody specs it and nobody reviews it. We do both. The AI removes the typing bottleneck. The senior engineer keeps the standards.
Spec first
Every feature gets a written spec before any code exists. The architecture, the data model, the edge cases, the auth boundaries. The spec is the contract.
Claude Code implements
Claude Code builds against the spec, driven by our own ecosystem of 20+ custom skills that handle scaffolding, tests, PR review, and security scans.
A senior engineer gates every line
An engineer with 10+ years reviews everything, with hard gates on auth, money handling, and security. The judgment stays human. Nothing ships unreviewed.
AWS-native, production-grade
The same serverless stack and standards we use for client work: Lambda, DynamoDB, EventBridge, Terraform, real test suites, real CI. A team’s six-month build, shipped in weeks.
Three products. In production. Built this way.
Not slides. Real products with real users, each shipped in days or weeks.
CostPatrol
AWS cost anomaly detection SaaSShipped in 2 weeksScans AWS accounts daily across 127 rules over 38+ services, flags anomalies within 24 hours, and delivers the exact fix command with a dollar amount to Slack. Serverless on Lambda, DynamoDB, SQS, and EventBridge. Stripe billing, CloudFormation onboarding in 5 minutes, read-only access.
VowTrust
Institutional deal verification platformWeeks, not the usual six monthsDual-mode verification for banks, lawyers, and SMEs. Express runs six parallel compliance checks across five providers in 15 to 45 seconds and scores 0 to 100. Institutional puts human mediators in role-based, invite-only deal rooms. Strict lifecycle state machine, immutable audit trail, exportable certificates. Next.js, Lambda, PostgreSQL, Terraform.
Boody
AI nutrition coach on WhatsApp and TelegramConcept to live users in 3 daysA self-hosted Gemma 4 26B vision model reads food photos and returns calories, protein, and a normalized name. Ollama on local Apple Silicon for the pilot, vLLM on cloud GPUs in production. A 4-layer memory system (SQLite, LanceDB embeddings, weekly AI coach notes, recent context) makes it feel like a real coach. Roughly $0.001 per interaction versus $0.01 to $0.05 for API-based competitors.
What this does not prove
- The repositories behind these numbers are private, so the figures are self-measured and dated rather than independently verifiable. The live products are the part you can check yourself.
- A commit count measures activity, not judgement. It is here because the revert rate next to it means something, not because volume is an achievement.
- The pipeline catches classes of error it has been taught to look for. It did not catch the ones in this paragraph; a person did.
All figures measured 2026-08-10.
Questions a skeptic asks.
Yes. Claude Code implements against a written spec, and a senior engineer reviews every line with hard gates on auth, money, and security. The AI does the typing. The architecture and judgment stay human. This is how CostPatrol, VowTrust, and Boody were built.
Still have questions? Book a call