How we work
- Sheet
- A-500
- Set
- Practice
- Status
- Issued
Multiple models. Independent review. Human ownership.
Defined AI roles, isolated tasks, provider‑enforced spending caps and independently reviewed changes: a structured engineering process behind every build, from a two-day assessment to a product that runs for years.
A-500PracticeIssued
-
01
Discover
Engineer
We read the system before touching it: the code, the data model, what is actually deployed, and what the business depends on. On an AI‑built app this includes a pass for the patterns those apps repeat, like exposed secrets, missing row‑level security, and sessions that are never checked.
-
02
Architect
Human decides
We decide what to keep, what to change and what to replace, and we write down why. A rewrite is a decision, not a default. When a fix is small and the foundation is sound, we say so.
-
03
Build
AI, human-directed
Engineers direct the work and AI speeds up bounded pieces of it. Our internal system runs scoped tasks across several models, each in an isolated environment with a hard spending limit. It never touches your production data.
-
04
Attack / review
Second seat
A second, independent seat reviews every change before it counts as done. The model that wrote the change never grades its own work. Reviewers are told the domain and the stakes, because a finding that matters for a $50,000 invoice can be noise on a $50 one.
-
05
Verify
Human decides
We reproduce a finding before acting on it: write the failing test, confirm that it fails, then fix it. If the test passes, the finding was wrong, and the test stays as evidence either way.
-
06
Ship
Human decides
You get the diff, the tests, and a plain account of what changed and why, not just a green checkmark. Deploys go through the same environments and controls we use on our own systems.
-
07
Observe
Engineer
When an engagement continues past delivery, we watch what we shipped: errors, performance, and whether the fix held. A control nobody checks isn't a control.
Every change gets challenged before it counts.
Where AI does the work
Bounded, well-scoped implementation: a function, a test suite, a migration, a documented fix. Each task runs in its own isolated environment under a hard spending limit, and never touches production systems or credentials.
Where a human decides
Architecture and scope. What we keep, change or block. Whether a review finding is real. What ships. Nothing ships on AI review alone.
Engagement models
Fixed scope for defined work. A bounded look first when it isn't defined yet.
- Defined work
- Fixed price and fixed scope. You know what you'll pay before you pay it.
- Unknown size
- A short, priced discovery or assessment comes first. We never bill hourly on a problem nobody has sized.
- Access
- You get access to the repository and CI from day one, and the work lives in your systems, not ours.
- Handover
- Documentation, decision records and a walkthrough are deliverables, so the knowledge doesn't leave with us.
See this applied to your system.
Start with a project brief, or read how the prototype-to-production assessment works.
Next sheet · A‑600