One engineer, many agents.

Akribia is my workshop, not an agency. The content flood around “building with AI” is hype-merchants on one side and doomers on the other; the honest middle — spoken by someone with production scars — is nearly empty. That’s the gap this publication fills: what building real software with agents actually looks like — complete systems with backends, deployments and monitoring, not another frontend demo. Every build log ends with a verdict: what worked, what it cost, what blocked.

Everything here is shown. If I claim a method works, there’s a PR behind the claim. If it didn’t work, that gets published too — the roadblock entries exist because nobody selling something can afford to write them.

The opposite of vibecoding. What survives review, in the order it happens:

• Spec before code. Every change starts as a written plan an agent (or I) can be held to — scope, contract, verification criteria.
• Adversarial review. Multiple reviewer agents, each briefed to refute the work and each other — not to agree. This has caught bugs I would have shipped.
• Contract-first codegen. One schema as the single source of truth, with every language binding generated from it, so interfaces can’t drift.
• Isolation. Parallel agent work happens in isolated worktrees; nothing lands unmerged, unreviewed, or untested.
• Verification before “done”. A change is finished when its behaviour is observed, not when the diff looks right.

The full method, applied end-to-end, is documented in the QuantScreen case study and taught in the course.

The brand is the publication; the voice is the engineer. The work is public — the PRs, the failures, the numbers — and that’s where the credibility has to come from, not from a headshot. My legal name sits where German law requires it: the imprint.

QuantScreen, my quant equity screener and the flagship case study. Mage Survivor, a bullet-heaven incremental RPG built with the same discipline. And the newsletter — AI & Tech, distilled into one considered digest.