Articles
Real-world engineering case studies about building software with coding agents: what I delegate, where agents fail, how I verify their work, and what remains the engineer’s responsibility.
Original articles in Russian: Статьи.
From STL to BREP: How an AI Agent Learned to Recover Geometry, Not Just Stitch Triangles
The convert_mesh_to_brep tool for FreeCAD MCP recovers planes and cylinders from STL instead of leaving faceted surfaces: 2884 triangles → 31 faces. Results on six parts, the limits, and a request for an STL that breaks it.
When an AI Agent Stopped Just Writing Code and Started Testing the Whole System
ESP32, an Android tablet and a Kotlin BFF became a single test bench: controlled Bluetooth failures, recording recovery and uploads to the server. Long runs, discovered bugs and the limits of the results.
The agent could move parts but couldn’t assemble a mechanism
How I extended FreeCAD MCP after an hour and a half of assembling through chat: Claude and Codex each assembled the pan-tilt on their own, yet the geometry checks missed the swapped side plates. An analysis based on logs, models and measurements.
Don’t write the prompt. Get interviewed.
I asked Codex to interview me instead of writing a spec: an hour of questions about a phone-based document camera, then the agent built the app and wired it into PaperLedger.
My Git Flow, Part 3: I Stopped Painting Cells in Confluence, the Release Stage Is Now Computed by a Portal and an MCP
How a release stage is computed from merged merge requests: a portal instead of a Confluence table, and an MCP tool that tells a coding agent which branch to start a task from. And where this algorithm can be wrong.
My Git Flow, Part 2: I Got Tired of Policing the Process and Handed It to GitLab
How a versioned stage/version branch name became the basis for merge request checks, build and environment selection, delivery, release notes, notifications and Jira statuses, and what still stays with people.
GitFlow as Cement Between Planning and Development
The GitFlow variant I have used for more than ten years: every release has its own chain of stage/version branches, a new version starts from the current stage of the newest one, and fixes move forward. Branches mirror the delivery plan, so release scope is not reassembled with cherry-pick and revert.
I Wanted to Read Three Columns of an Invoice
OCR turned out to be the easy part: PaddleOCR running out of memory on a real phone photo, aligning by the table grid instead of the sheet of paper (a hint from an old Sudoku project), and a tight crop instead of a new model.
The Feature I Wouldn’t Have Touched Myself
How a coding agent added Jetpack Compose to Android UI Renderer MCP: it parsed Kotlin signatures itself, built a Compose probe and a semantics tree, while I set the boundaries of the solution and caught errors a single PNG couldn’t show.
The android-ui-renderer-mcp project
One Layout Wasn’t Enough: Rendering a Whole Android Screen Without Running the App
How a coding agent got a composed Android screen — an Activity with a toolbar, a Fragment, RecyclerView rows and a loading overlay — rendered from real XML with explicit data, plus a replay recipe for every render.
The android-ui-renderer-mcp project
What a Coding Agent Taught Me About A/B Test Telemetry
How a coding agent designed the analytics for an A/B/C experiment with a new Android price-tag scanner: a scan-session event model, the Firebase to BigQuery export, and a one-row-per-session layer it queries for me.
One Master Prompt, Five Steps: How a Coding Agent Built a 4-Platform Kotlin App
The real prompts behind a self-hosted Kotlin Multiplatform messenger: an architecture-contract master prompt and five sequential steps, from the shared contract to a self-review. All of them are public.
From a Paper Map to a Problem Split
How I combined a scan of an orienteering map, a GPS track, and an official race protocol to find recurring mistakes on specific legs of a course.
I Gave a Coding Agent Eyes for Android UI — but an Image Was Not Enough
Why a screenshot of an Android interface was only the first step for a coding agent. And how a View tree made it possible to check layout geometry without manually reviewing every result.
A Coding Agent Almost Fixed My CameraX Bug with a Delay. Why I Stopped It
How we investigated a rotation mismatch between Preview and ImageAnalysis on an Android tablet. And why we stopped the first, seemingly workable fix that relied on a delay.
I Sped Up Development with AI Agents. Testing Became the Bottleneck
Why an AI agent needs an MCP for Allure TestOps when a human can just open a browser — and how the agent started carrying meaning between code, the TMS, and MockServer.
The allure-testops-mcp project
I Didn’t Need Another Gas Station Tracker
How a personal family portal stopped being a fuel monitor and became an assistant that decides “go for 95-octane or wait” — and stays quiet until it matters.
I Wanted AI to Improve My Resume
Why “update my resume with AI” turned into a private fact base about my own career — and how the agent became more useful once it stopped writing text and started asking questions about metrics.
How I Turned AI into a Personal Coach
Why screenshots from Mi Fitness stopped being enough after a run — and how a personal portal gave an AI coach the training history, recovery metrics, and how-I-feel context it needed.