Manifesto
I run a production Kubernetes cluster. With real customers. I cost it, I steer it, I defend it in front of them. And when it breaks at three in the morning, I'm in the logs.
6
nodes
RKE2 v1.32 on OpenStack, run by a team of five.
~60
applications
continuously reconciled by ArgoCD.
4
customers
hosted in production and staging.
90%
less storage
486 → 48 GiB of backups after one post-mortem.
Selected work
From infrastructure to production, from code to consulting. Keep scrolling — each project has its own page.
Product Owner · team of 5 · Cassiopée GitOps portal
The SRE team running the major's production Kubernetes cluster. I own the backlog, decide what ships — and I'm in the logs when it breaks. With Cassiopée, a complete environment deploys from a single commit, with no platform team in the loop.
Azure × private cloud
A self-service platform that stays up when either cloud goes down. Azure carries the front ends, OpenStack carries the state, a WireGuard tunnel joins them and Patroni fails PostgreSQL over with no human involved. Everything is described in OpenTofu: rebuildable from scratch.
Repository-first
An internal platform where the standard path is also the fastest one. A GitHub App installs the pipeline in the repository, publishes the image to GHCR and deploys to AKS. A team goes from empty repo to first deployment without writing a line of pipeline.
Global BI · Sept 2025 – Jan 2026
The Excel export engine of a finance portal used in 90+ countries crashed past 50,000 rows. Rewritten as pure streaming — Open XML SDK in SAX mode, sequential reads, direct writes to disk: 1 million rows in ~2 minutes.
DevOps workstream lead · 50 participants
The information system of a miniature airline, built by 50 participants. I led the shared foundation: API conventions, GitLab CI/CD, Prometheus / Grafana / Loki observability and a RabbitMQ bus. A standard only gets adopted if following it is faster than working around it.
Target architecture
Scoping a Finance transformation on SAP S/4HANA: conceptual foundation of the domain, mapping of financial flows and interfaces, then a target architecture deck defended before a panel — to business stakeholders as much as technical ones.
Oracle HCM · Monte-Carlo estimation
HRIS transformation on Oracle HCM. Instead of a single number pulled from a spreadsheet, an effort estimate by Monte-Carlo simulation at P80 — a defensible range with its confidence level — plus the workstream's risk register.
Security
A full EBIOS Risk Manager assessment on a target information system, from scoping to the treatment plan. The method forces you to name an attacker, their objective and their path — and to keep the link between the real threat and the technical control.
Co-founded B2B startup
A connected locker that uses AI to check equipment condition on return. The first market, schools, couldn't pay: we repositioned towards industrial field service. I built the hardware BOM, the 5-year financial model and the investor pitch.
Product ownership
Running a platform also means deciding what to build, for whom and at what cost. Five principles — each one held on a real project.
01 discovery
Co-founded B2B startup: schools wanted the product but couldn't pay for it. We repositioned towards industrial field service — with the hardware BOM, the 5-year financial model and the investor pitch to back it.
View project · Lendr02 backlog
Product Owner of a five-person SRE team: I own the backlog and decide what ships. Even a SPOF found during an audit goes in with its exit options, not as an orphan alert.
View project · SIOPS03 estimate
Instead of a single figure pulled from a spreadsheet, an effort estimate by Monte-Carlo simulation at P80 — a defensible range — plus the risk register that goes with it.
View project · BeBlood04 adoption
A shared foundation for 50 participants: API conventions, CI/CD, observability. Same logic on CNP: a team goes from empty repo to first deployment without writing a line of pipeline.
View project · UBSI05 alignment
Four customers hosted on SIOPS, a SAP S/4HANA target architecture defended before a panel: turning a technical choice into cost, risk and lead time for business stakeholders.
View project · DanonePost-mortems
Three incidents on the SIOPS cluster. Symptom, diagnosis, fix, lesson — unfiltered.
rke2-server waits on kubelet, which waits on etcd. Deadlock.rke2-killall to break the loop, clean restart, then resize the master.forget --group-by host + prune: 486 → 48 GiB. Script hardened with set -euo pipefail.Career
Product Owner — SIOPS
Five-person DevOps/SRE team, production RKE2 cluster on OpenStack.
Ipsos — Software Developer, Global BI
Rebuilt the Excel export engine of a finance portal used in 90+ countries.
FPT University, Vietnam
Exchange semester. TOEIC 900.
EPITA — SIGL major
Information Systems & Software Engineering. Class representative.
Clayens NP — operator
First contact with industrial production and its constraints.
Scouting — 4 years as unit leader
Leadership, logistics and responsibility for a group in the field.
Stack
No percentages. The number next to a tool is how many projects I used it in — hover to see which.
Infra & Cloud
GitOps & CI/CD
Security & secrets
Observability
Development
IS & consulting
Daily driver: Arch Linux.
Contact · reply within 24 h
6-month pre-hire internship from February 2027 — Platform, Cloud, SRE or DevOps, open to technical Product roles. Email lands in my inbox, not in a form.