lapetina.fr

AI governance / architecture models / developer systems / applied AI

David Lapetina

I work where AI agents meet real systems: Syncalytics for enterprise AI governance, and three open source projects — YaC4e for architecture as a living model, DevDeck for leaner AI-assisted coding, Leading Tone for teaching music theory from a computed model.

Syncalytics YaC4e DevDeck Leading Tone
Focus
Governed AI operations
Open source
YaC4e, DevDeck, Leading Tone

Four projects

Different surfaces, one idea: AI should operate on explicit structure, under human decisions.

Syncalytics.com

AI governance for production agents.

Syncalytics is an enterprise AI governance platform for teams moving agentic AI from pilots into operational systems. It focuses on agent identity, runtime guardrails, approvals, budgets, lineage, and audit-ready evidence.

  • Control prompt, tool, retrieval, memory, MCP, and agent-to-agent events while they happen.
  • Connect policy decisions to approvals, traces, lineage, and reviewable evidence.
  • Support regulated, sensitive, and domain-heavy AI workflows with enforceable controls.

Commercial platform · syncalytics.com

Explore Syncalytics
YaC4e

Architecture as a living model, not a drawing.

YaC4e — "yet another C4 editor" — is an open source web application for maintaining a metadata-rich C4 architecture repository. Elements, relationships, metadata, and views are stored as canonical data, so each diagram is a projection of one shared model rather than a separate picture.

  • Model C4 elements and typed relationships with ownership, lifecycle, security, and custom metadata.
  • Import Structurizr DSL, export canonical JSON, YAML, SVG, and PNG, and validate governance gaps.
  • Give agents a read API for context and impact analysis, and keep their proposals behind human review.

Open source · MIT · Java 21, Quarkus, React 19

View YaC4e on GitHub
DevDeck

A command deck for AI-assisted coding.

DevDeck is an open source project to make coding with AI lighter and more deliberate: keep useful commands, project actions, and workflow context close without relying on heavy IDE surfaces such as IntelliJ or Visual Studio Code.

  • Keep repeatable development commands visible, searchable, and easier to run.
  • Reduce tool switching when working with terminal-first AI coding agents.
  • Favor a small, inspectable workflow layer over a fully loaded editor environment.

Open source · developer workflow

View DevDeck on GitHub
Leading Tone

A tutor that decides in code, and only borrows the words.

Leading Tone is an open source desktop tutor for music theory, from notes through to species counterpoint. It keeps a model of what you actually know, chooses what to teach from it, sets the question, and marks the answer — typed, or played on a MIDI keyboard.

  • Thirty-six concepts with prerequisites instead of a syllabus, so the route through them follows the evidence you have given.
  • Every decision — concept, question, verdict, mastery — is computed and tested; the language model supplies wording only, and switching it off changes nothing else.
  • Explanations cite published sources and show real annotated bars from score corpora, or say plainly that no example exists.

Open source · MIT · Desktop app · Java 21, Quarkus, React 19

View Leading Tone on GitHub

Working thesis

AI systems need a control plane and a source of truth. Developers need less ceremony.

My work sits at the point where AI leaves demos and enters real operations: identity, permissions, policy gates, evidence, and human decisions have to be part of the execution path. That is what Syncalytics enforces at runtime.

Agents also need something dependable to reason about. YaC4e turns architecture into canonical, queryable data with an agent-facing read API, and routes proposed changes through a human review step instead of letting them land silently.

On the developer side, AI coding should make the local workflow sharper. DevDeck is my way to keep commands and context accessible without turning every task into a heavyweight IDE session.

Leading Tone takes the same rule into a domain where it is easy to check: what to teach, what to ask, whether the answer was right, and what the learner now knows are all computed and covered by tests. The model is handed a decision and asked for the words, never for the verdict.