ENGINEERING PRACTICE · CASE STUDY

Building with AI

We Stopped Writing Most of Our Code. Delivery Got Better.

Industry

Software Engineering

Services

AI-Assisted Delivery, Code Review, QA Automation

Platform

Internal · All Products

Timeline

Ongoing

Year

2026

Building with AI product screenshot

THE STORY

From Typing Code to Directing It

The honest version is that we were sceptical. AI wrote plausible code that fell apart against a real codebase, and reviewing it took longer than writing it. What changed was not the models getting better — it was us treating AI as an engineer that needs context rather than a search box that returns snippets. We started giving it the same things a new hire gets: the architecture, the conventions, the reasons behind past decisions, and the tests that define correct. Then we stopped asking for functions and started asking for changes, reviewed the diff the way we review anyone's, and let the test suite be the argument-ender. The engineers did not write less software. They stopped spending their day on the parts of it that were never the hard bit.

Quick Facts

Client

Techeffic

Team Size

Every engineering team

Duration

Ongoing

Technologies

Claude, GitHub Actions, Codemagic, .NET, Angular, Flutter

The Bottlenecks

The Challenge

TRUST

Code You Did Not Write

Reviewing generated code properly is slower than skimming it, and skimming is how bad code gets merged. Nothing ships without a human review and a passing suite — the same bar as any other change.

CONTEXT

A Model With No Memory of Your Codebase

Generic answers come from generic context. The work was writing down the conventions, the architecture and the reasoning that lived only in people's heads — which turned out to help the humans just as much.

SECURITY

Client Code Leaving the Building

Client work carries obligations. What can be sent to a model, what has to stay inside, and how secrets are kept out of a prompt are decided per project before any of it starts.

Look Book

Product Interface Showcase

Product interface screenshot
Product interface screenshotProduct interface screenshot
Core Modules

Key Features

🧭

Specs Before Prompts

Work starts as a written change with acceptance criteria. A vague request produces vague code whoever writes it.

🔍

Review Is Still Human

Every diff is read by an engineer who has to defend it. Generated code gets no shortcut through review.

🧪

Tests Are the Contract

Behaviour is pinned by tests before the change, so correctness is measured rather than argued about.

📓

Decisions Written Down

Why a thing is built the way it is now lives in the repo — the same context a model needs is the context a new engineer needs.

← Back to all case studies