I help teams take GenAI from architecture to production, with measurable quality, cost and governance. Based in Paris, I am the founder of Digitalizor and an external Chief AI Officer (CAIO), with 20 years of experience across IT, data and AI.
My focus: observable AI systems, evidence-based governance and controlled agent workflows.
Website · LinkedIn · Malt · Technical portfolio
| Project | Purpose | Availability |
|---|---|---|
| Grafana LLMOps Forge | Discover existing telemetry and generate Grafana dashboards for LLM costs, agents, quality and governance. Automated query/deployment checks; separate visual review. | v2.0.4 |
| OpenShadow AI | Correlate network, endpoint and identity signals for self-hosted Shadow AI discovery. Each finding distinguishes presence from observed usage. | v0.3.0 preview; controlled pilots |
| FTP Deploy MCP | FTP/FTPS/SFTP deployment with root confinement and host-key pinning. Confirmation, staged promotion and rollback belong to the optional durable workflow (stateDir). |
v0.3.2; npm + MCP Registry |
| OpenSpanGuard | Enrich OTLP/HTTP JSON generation spans with evaluation signals and coverage, preserving trace identity and timestamps. | Developer preview; source + offline demo |
| TierDecay | Certify cheaper model tiers for recurring coding tasks, with statistical checks and recertification when models change. | v0.5.0; CLI + MCP |
| Claude, Inc. | Structured project workflows with departmental skill manuals, bounded reviews and artifact records. | v1.5.1: 54 manuals / 8 departments. Development source: 64 / 9, with CAIO and Growth. |
- Forge: Generated PromQL queries executed against Prometheus and Grafana E2E checks. Semantic visual review is a separate human/agent step.
- TierDecay: controlled pilot and protocol replay: 36 executor runs + 3 distillations; 24% estimated savings in replay versus always-frontier. The 0.657× result compares the fast tier without a playbook with frontier, using normalized warm-cache costs. Both results are bounded to this fixture.
- FTP Deploy MCP: 43 scripted scenarios in each of English and French, using a real MCP SDK client and an in-memory remote adapter. The repository documents transport and client qualification separately.
harnessmeter estimates instruction/context token footprints. npm 0.3.0 covers the published Claude workflow; Codex, static audit, baselines and CI budgets are currently available from unreleased source. Local profiling is offline by default; optional T2 review invokes a model provider.
Generative Engine Monitor compares brand visibility in provider API responses, with confidence intervals and replay from stored responses. dsh-plugin-otel-genai provides content-free telemetry instrumentation. awesome-ai-architect maps the GenAI engineering stack.
- AgentShield #123 — merged: root
.mcp.jsondiscovery/classification and security regression coverage. - LibreChat #15046 — merged documentation: portable agent-source paths and
CLAUDE.md/AGENTS.mdguidance parity. - Claude SEO #285: JSON-LD detection fixes, cherry-picked into v2.3.0 with attribution.
- Grafana plugin-tools — merged: signed-commit guidance and Node 24 scaffold.
Ongoing proposals: OpenTelemetry GenAI #516, under review; OpenWiki #818, draft, awaiting live-provider validation.
Architecture and RAG · evaluation, LLMOps and FinOps · AI inventories and audit evidence · controlled tool access · external CAIO and team training.
I work with EU AI Act, ISO/IEC 42001 and NIST AI RMF evidence requirements. My working principles are explicit failure modes, least privilege, reviewable components and measured outcomes.
Core stack: Python · TypeScript · MCP · OpenTelemetry · Prometheus · Grafana · Docker.
Paris, France · French, English and Spanish · Contact
Project and contribution status reviewed 22 October 2026. Releases and development source are distinguished above.