ABOUT GEARJUNKIE
Learning AI should
feel achievable.
gearjunkie is a learning resource by Marcel Piontek for developers who want to use AI at work. We explain one idea at a time, without assuming you already know the AI landscape.
01 / UNDERSTAND
One concept, one example.
Short explanations with simple technical examples. Unfamiliar terms are explained in the text and glossary.
02 / TRY
From reading to doing.
Each basic includes a small exercise. A one-minute short explains the core idea with animated graphics, narration and on-screen text.
03 / CHECK
Sources instead of promises.
Facts must trace to official documentation or the original project. Community discussion helps discover topics but cannot replace evidence.
FOR LEADERS & BUILDERS
Why this matters to CEOs and IT colleagues
CEOs & leaders
If you want to assess AI projects or assemble a team to build one, you need a sense of what sits behind an automated application. gearjunkie makes concrete questions tangible: Which tasks belong to the model? Which rules does software check? Where do costs arise, and when may a result be published?
The project shows how responsibilities, traceability and budget limits can be built into a workflow. It provides a working example for assessing your own AI initiatives. A financial benefit for your business would need to be evaluated against your own tasks and quality criteria.
Developers & IT colleagues
You can start with the basics and use the underlying system to see how AI fits into an application. For development, architecture and operations, validated data hand-offs, stored run records and explicit failure paths are particularly relevant. Product and project teams can see how a technical capability becomes an understandable learning resource.
A practical starting point: choose a lesson, try the exercise and consider which checks you would need for the same use in your team. This connects foundational knowledge with implementation.
THE WORKFLOW
A digital editorial team with clear roles
Several specialized AI roles handle bounded tasks in sequence. They exchange structured data. Additional checks in application code help decide whether the workflow may continue.
Scout
Discovers suitable basics and gathers sources. Community discussion can suggest topics; factual claims require registered primary sources such as official documentation or the original project.
Concept Analyst
Builds a concept dossier with supported claims and sources. This provides the foundation for the explanation.
Editor
Writes the English lesson, German version, example, exercise, glossary and scenes for both shorts.
Publisher
Reviews the lesson and short together and provides feedback. Specific issues are repaired within a bounded number of attempts. If issues remain, publication is blocked.
Both language versions of the shorts are generated and checked before publication. Narration comes from approved scenes; a separate transcription checks the spoken content. The new lesson is written only after the text, evidence and both shorts pass the required checks. Public topic suggestions never start a paid run: an operator must explicitly initiate one.
CONCEPTS IN PRACTICE
The concepts used here
These AI and software concepts describe how gearjunkie is implemented. Each explanation connects the technical name to its specific use in this project.
Multi-Agent Architecture
Several AI roles divide research, analysis, writing and review. Each role receives its own task. This makes individual steps and their feedback traceable; a reviewing agent can still be wrong.
Grounded Generation
The editorial team receives claims with associated sources. Facts must trace to registered, reachable primary sources. This reduces unsupported claims but guarantees neither correct interpretation nor error-free text.
Structured Outputs
Agents exchange data in a defined format. Schemas check that required fields and sources are present, for example. Application code can then process results and reject incomplete hand-offs.
Deterministic Guardrails
These are fixed checks in application code. They validate source trust, safe URLs, glossary links and publication conditions, among other things. Model approval cannot override them. Language quality is also reviewed by AI.
Persistent State
SQLite stores publications, covered concepts, run records, costs and feedback. Files preserve intermediate outputs and review reports. This stored state helps prevent duplicate publications and reconstruct a run later; it does not train the model.
Model Routing & Budget Controls
Model assignments are configurable per role. Costs are recorded per run, and paid steps are subject to a monthly budget limit. Local estimates are not provider invoices. Fixture runs using fixed example data incur no API costs.
AI-as-a-Judge & Evaluation
AI reviews text and compares proposed improvements. A blind comparison randomly assigns A/B labels so the evaluator does not know which version is new. Multiple cases and fixed checks supplement the model judgment; it remains a fallible assessment.
Controlled Self-Optimization
A separate Optimizer may propose exactly one bounded change: an additional Editor instruction or an allowed Editor model. Activation requires passing quality and cost evaluations. Configurations remain versioned and reversible. It cannot alter code, source rules or protective checks.
CONTROLLED IMPROVEMENT
Improvement through a verifiable process
The Optimizer examines run outcomes, issues, costs and feedback. A proposed change must then compete against the existing configuration on at least three representative cases. Blind comparison and fixed checks must meet the configured quality and cost criteria. Without demonstrated improvement, the existing configuration remains. The system does not invent its own publication rules.
UNDER THE HOOD
What you can inspect technically
The website uses Kirby and PHP. A separate TypeScript worker orchestrates the editorial workflow, SQLite holds state, Docker provides the runtime environment and systemd schedules recurring runs. Animated shorts use SVG graphics and FFmpeg; AI generates and verifies the voices. The repository documents roles, schemas, source rules and the controlled optimization process. You can trace the architecture down to its concrete checks.
AI-generated, openly disclosed
You can suggest topics and report mistakes to marcel@gearjunkie.de. The previous technical edition remains in the archive.
Open the learning path ↗