In-depth analysis · About 6 min read
What three Java Champions reveal about verifying AI coding work
GitHub’s Java interviews show how context, platform reproduction and explicit checks turn Copilot output into reviewable engineering work.
- Published here
- Source published
- Source checked
GitHub published interviews with Vlad Mihalcea, Patrick Baumgartner and Frank Delporte on October 20. Their examples span feature development, teaching materials and a silent hardware-input bug. The useful common thread is a verification process: developers define the problem, inspect evidence and retain responsibility. These are practitioner accounts published by the tool’s vendor, not a controlled productivity study.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- Baumgartner reports migrating about 900 slides across 15 presentations in eight to ten hours, using screenshots, animation inspection and PDF comparisons to check the result.
- Delporte describes a Pi4J V4 failure in which the fifth input watcher could not run after native polling occupied the four carrier threads in the reported Raspberry Pi setup.
Give the assistant the project’s existing standards
Mihalcea uses Copilot inside IntelliJ IDEA for investigation, debugging and prototypes. He says an Optimizer event-history web application, repeatedly postponed because he expected weeks of work, took two days to implement for a forthcoming Hypersistence Optimizer version. That is his comparison with an anticipated effort, not a timed experiment against another implementation. He still reviews plans and generated changes when the answer is unsuitable.
His reusable skills draw on 65 published articles and 12 years of blogging to shape documentation and preserve expressive, testable, simple code. He describes choosing Opus with a one-million-token context and Extra High or Max reasoning for difficult work, and lighter models for reports or documentation checks. This describes his workflow, not proof that every Copilot plan offers those settings. Our interpretation is that project conventions and task sizing matter alongside model choice.
A migration needs evidence that meaning survived
Baumgartner’s Reveal.js-to-Marp migration covered roughly 900 slides in 15 presentations, including code, animations and live JavaScript examples. Before execution, he and Copilot analysed the material, specified the target, constraints and examples, then moved from structure to details. He reports eight to ten hours of work, compared with his estimate of a week or more manually.
The acceptance evidence was broader than converted files: screenshots were compared visually and algorithmically, both decks ran side by side for manual animation checks, and generated PDFs were compared. Typos could be fixed automatically, but meaning, versions, dates and content changes required discussion. For AISubHub, this illustrates why file coverage and successful conversion are insufficient when the actual requirement is preserving a teaching sequence.
Reproduce the user’s platform before declaring a fix
A terminal application built with Tamboui worked under Linux in WSL, yet a user found that Enter was ignored in some forms in native Windows CMD or PowerShell. Baumgartner describes turning that report into UI tests, using tmux on Linux and discovering psmux for similar PowerShell flows. They reproduced the behaviour, created a local fix, traced the fault into Tamboui and submitted an upstream change that was merged.
When test descriptions exceeded the available context, they replaced repeated tables with a compact sequence notation while retaining critical information. The lesson is not to shorten away assertions. Our analysis is that a representation should preserve actions, input and expected behaviour; a Linux-only pass cannot settle a Windows report. The interview does not supply a universal compatibility matrix or exact runtime versions.
Resolve requirements before accelerating execution
An educational app changed direction after discussion of learning methods and actual curricula: rather than immediately serving every country, Baumgartner focused on Switzerland. It covers arithmetic for grades one through four, offers German, French, Italian and English, provides explanations and repeated practice after mistakes, and includes two-player modes. He chose no backend, keeping children’s data on the device and hosting the app on GitHub Pages. This is a design choice described in the interview, not a security audit.
On a roughly four-hour journey, he also steered work on a Maven plugin for Spring Boot auto-configuration from his phone, reviewing a CI pipeline, sample-application performance tests and reports. An earlier requirement to support both Spring Boot 3.5 and 4.x produced two architectural paths; later removing the unneeded 3.5 path was costly. These examples support small decisions and early clarification, not a promise that any plugin can be delivered or validated in four hours.
Why a fifth button could stay silent
Delporte’s example concerns Pi4J V4, released in February 2026 with a Foreign Function & Memory API plugin. Four input buttons worked while the fifth produced no event, exception or warning; the older gpiod plugin worked on the same hardware and application. He gave Copilot two starting paths, the provider class and FFMDigitalInput, and reports receiving a proposed pull request about half an hour later.
In the reported setup, each watcher blocked in native poll() while waiting for a GPIO change. A virtual thread blocked in that native call stayed attached to its carrier. Four watchers occupied the four carriers on the four-core Raspberry Pi 4, leaving the fifth unscheduled. That causal explanation is more informative than calling events unreliable. It remains the interview’s diagnosis: it does not establish that every Java version or virtual-thread operation behaves this way. Maintainers reviewed and approved changes before merging.
Turn lessons into checks, not an unchecked promise
Baumgartner describes reusable skills such as maintainer-audit and mentions unlazy as an example of explicit completion gates. His quality harness combines Checkstyle, PMD and Spring Modulith architecture checks so the agent can change, run checks, read findings and iterate. He selects VS Code for visual feedback, terminal or coding-agent workflows for larger delegated work, and MCP integrations only when useful; a good CLI may be preferable. These are his practices, not skills tested by AISubHub.
Delporte adds a strong diagnostic check: a regression test should fail on the old code, then pass after the fix. He also asks for an understandable explanation and reviews duplication and maintainability. Our recommended adaptation is to define task-specific evidence before execution, check it independently and retain human acceptance. Static checks alone cannot validate animations, native platform behaviour or an educational product’s suitability; the original accounts still require people to run and judge the software.
Why it matters
AISubHub analysis: the transferable unit is the complete feedback loop, including failed assumptions and checks. Teams can pilot one representative task, record its environment and inspect whether the proposed evidence would expose a wrong answer before expanding delegation.
Limits and availability
The time estimates and outcomes are interviewees’ reports. We have not run their code, reproduced the hardware issue, inspected the linked pull requests or audited their skills. The article gives no controlled baseline, general speedup or universal model entitlement; runtime and platform details must be checked for each project.
Sources and evidence
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.