Baller Knowledge UK · Koachie

AI Code Assurance

Generated code, verified before it ships. One plain-English sentence becomes a complete feature — database schema, server function, authentication, front end — and every piece is gated by checks that cannot hallucinate, because they are rules rather than a model. The generator is open-source Qwen2.5-Coder-7B; the checks around it are ours.

> a website for tarot readings that takes appointments

A real recorded build

Watch it work

This is a genuine run, captured event by event on a single desktop GPU. Nothing here is mocked — the code, the checks and the repair below are exactly what the pipeline produced.

runs in about 3–4 minutes on an RTX 3070

Pipeline

waiting

Checks

· not started
Press play to watch a real build.

Why it is different

The model is allowed to be wrong

Every AI writes code. The difference is what happens next. Four independent gates stand between the model and you, and each exists because of a failure we measured rather than imagined.

Relevance gate

Scores how well the library actually matched before generating. When nothing relevant was found it says so, instead of answering from memory behind invented citations.

Caught a booking request retrieving Docker CPU statistics code — and the earlier pipeline citing it as a source.

Decomposition

Splits the request into separate artifacts and builds them one at a time, so fixing one cannot break another.

All at once: 4/6, then 3/6, then 3/6 across correction rounds. Decomposed: 10/10.

Security guard

Deterministic rules inspect the generated code and block it on violation — password columns, SQL injection, exposed admin keys, tables with no access control. Each rule cites its source.

Caught 5 of 5 issues in the generator's own insecure output.

Self-repair

Some fixes are mechanical, with exactly one correct form. Those are repaired automatically and always reported — never silently.

A schema arrived without row-level security twice, despite being asked twice. Now a rule fixes it every time.

What it catches

The mistakes that cause real breaches

ProblemWhy it mattersAction
Password column in an app table Duplicates a secret the auth system already protects properly block
Table without row-level security Every table is reachable by API — without it, anyone reads everything repair
SQL built by string concatenation The most exploited web vulnerability there is block
Admin key in browser code Bypasses every access control at once block
Policy clause inside a table definition Invalid SQL — the schema will not even run block
Deprecated authentication method Removed from the current API; fails against a live project warn
Request input used unvalidated How bad data and injection get in warn

What is ours, and what is not

We did not train the model. That is the point.

The code above was written by Qwen2.5-Coder-7B, an open-source model released under Apache 2.0, running locally on one desktop graphics card. We did not train it and we do not claim to have.

Theirs — the generator

Qwen2.5-Coder-7B-Instruct, Apache-2.0, ~4.7GB in memory. Swappable: the pipeline holds no opinion about which model writes the code, and a better one can be dropped in tomorrow.

We also train our own models from scratch — a 312M with open weights and a published evaluation card — but it is not what writes this code, and saying otherwise would be a lie.

Ours — everything around it

The retrieval and its refusal gate, the decomposition, the security rules, the automatic repair, and the requirement checking. None of it is a model, so none of it can hallucinate.

Rules, not predictions. That is why they can be trusted to check something that can be wrong.

Why that is the stronger position

Model quality is a race between companies spending billions. Nobody wins it from a bedroom. But every model released makes our generator better and leaves our checks exactly as necessary.

Measured: a model 4.7× larger still wrote a password column and skipped access control. Size does not buy compliance.

Honest limits

What this does not do

Said plainly, because the alternative is being corrected in public

  • You cannot run it here. The pipeline needs a language model held in GPU memory and takes minutes per build, so this page replays a real recorded run rather than generating live. A live demo is available on request.
  • The guard is a floor, not a proof. Passing means the code did not match a known-bad pattern. It does not mean the code is correct.
  • Generated code still needs review. It is a strong, checked starting point — not something to deploy unread.
  • Verification is uneven by language. SQL and server code are rule-checked thoroughly; front-end markup much less so.
  • It is a working prototype, built and tested by one person, not a product with users behind it.