Alexandre Beguel, Agentic AI Architect & external CAIO

Open-source work

1. LLM observability and FinOps on an open stack

OpenTelemetry GenAI → Prometheus → Grafana, self-hosted, on the telemetry a platform already emits.

• grafana-llmops-forge

  Discovers what a Grafana deployment actually collects (OTel GenAI, LiteLLM, vLLM/TGI, DCGM), then generates, deploys and visually verifies seven dashboards for FinOps, agents, quality and governance, with multi-window burn-rate SLO alerts.

  v2.0 · zero dependencies · E2E CI on Grafana 9.5 → 13.2 · 63/63 generated queries executed on a real Prometheus · listed in awesome-observability

• openspanguard

  An OTLP enrichment gateway that evaluates completed generation spans, keeps original trace IDs and timestamps, and exposes evaluation coverage next to quality scores, so a good average cannot hide unevaluated traffic.

  developer preview · offline demo and Tempo/Alertmanager stack · project page

• dsh-plugin-otel-genai

  Content-free OpenTelemetry GenAI token and latency metrics for DeepSeek Harness, ready for the grafana-llmops-forge dashboards.

2. Evidence-first AI governance

• open-shadow-ai

  Self-hosted Shadow AI discovery that correlates DNS and proxy logs, Zeek/Suricata, an endpoint agent (processes, browser extensions) and Entra ID / Active Directory, with an explicit evidence level per finding: presence is never reported as usage.

  0.3 preview · 11-job CI with SBOM and signed image evidence · OIDC SSO and SCIM · EN/FR/zh-CN

• Governance blueprint of grafana-llmops-forge

  Maps EU AI Act articles 12, 26(6), 50 and 73, ISO/IEC 42001 and NIST AI RMF to live signals: logging continuity, automatic model inventory and incident watch. An evidence layer for auditors, not legal advice.

3. Agents under control

• ftp-deploy-mcp

  An MCP deployment server for FTP, FTPS and SFTP with local-root confinement, SFTP host-key pinning, SHA-256-verified staged promotion, explicit human confirmation and rollback.

  npm · Official MCP Registry · listed in awesome-mcp-servers · 43 scripted scenarios run in English and French against the real MCP SDK client

• tierdecay

  A per-repo learning layer for AI coding routers. It moves a recurring task class to a cheaper tier only after statistically confirmed successes, and recertifies it when models change.

  pilot: 0.66× frontier cost per task (95% interval 0.63–0.69), 24% saved against always-frontier · SBOM · OCI image · MCP server

• claude-inc

  A Claude Code plugin that runs a project as a virtual company, with peer CEO, CTO and CAIO, nine departments, 64 skill manuals, bounded review loops and SHA-256 records of every submitted artifact.

  v1.5 · the most-starred project of this profile · project page

• harnessmeter

  An offline context profiler for Claude Code and Codex that measures the token footprint of instructions and skills, enforces CI budgets against a baseline and proposes evidence-graded demotions.

  npm · zero dependencies · CI on Linux, macOS and Windows

Also: generative-engine-monitor measures brand visibility in AI answers, comparing parametric and web-grounded modes with bootstrap confidence intervals. awesome-ai-architect is a self-updating map of the GenAI engineering stack, classified daily by GitHub Actions (interactive explorer).

Upstream contributions

Merged and released

• LibreChat: merged. Made CLAUDE.md portable by replacing a machine-specific agents source path, and restored CLAUDE.md/AGENTS.md parity for auth-cache guidance.
• AgentShield: merged. The scanner now discovers and classifies root .mcp.json files, so MCP configurations reach the security rules, with regression coverage; all 16 review threads were resolved.
• Claude SEO: improved JSON-LD detection and removed template-expression false positives. The maintainer confirmed it landed in v2.3.0 with authorship preserved.
• Grafana plugin-tools: two merged contributions, signed-commit contributor guidance and a Node 24 .nvmrc scaffold update.

In review

• OpenTelemetry GenAI semantic conventions: recording token usage on fetch_response spans.
• LangChain OpenWiki: Mistral as a model provider.

Consulting & advisory

Engagements take GenAI from scoping to production for mid-sized companies, startups and e-commerce businesses. The repositories above show the method; each engagement applies it to the client's systems, data and constraints.

• Scoping and architecture: use cases, build vs buy, RAG, multi-agent systems (LangGraph), model and hosting choices.
• Measurable production: evaluation (RAGAS), observability, cost per team and per model, quality and latency SLOs.
• Governance and compliance: Shadow AI inventory, traceability and audit evidence against the EU AI Act, ISO/IEC 42001 and NIST AI RMF, controlled agent access to tools.
• External CAIO and training: AI roadmap, change management, upskilling teams on Claude Code and agentic systems.

FAQ

What does Alexandre Beguel specialize in?

Production GenAI and agentic systems. He covers the full path from scoping to operation: architecture, evaluation, observability, cost control and AI governance.

How can a team measure the cost of its LLMs?

Start from the telemetry already emitted, then attribute cost per team, model and provider. grafana-llmops-forge builds these FinOps dashboards in Grafana without a vendor SDK.

How can a company detect Shadow AI?

Correlate network, endpoint and identity signals, and keep presence separate from usage. open-shadow-ai does this self-hosted, with an evidence level per finding.

How do you prepare GenAI systems for the EU AI Act?

Build the evidence first: AI inventory, logging continuity, transparency notices and incident tracking. Article 50 transparency obligations apply since 2 August 2026. Under the AI Omnibus (Regulation (EU) 2026/1744), Annex III high-risk obligations apply from 2 December 2027.

Which languages does he work in?

French (native), English and Spanish. He is based in Paris.

Contact

Based in Paris, France. Works in French, English and Spanish.

• beguel.com
• LinkedIn
• Malt
• GitHub