Seven ways I work.
Priced per engagement,
not per hour.
Every engagement has a written scope, a date and a fixed quote. No rate card, because two companies with the same stack rarely have the same problem — and an hourly figure tells you nothing about what you get.
AI Sprint
Your company running on private, self-hosted AI in two weeks.
- Timeline
- 2 weeks
- Format
- Fixed scope
- Best for
- 20–200 people
- Runs on
- Your hardware
What you get
- OpenWebUI + Ollama deployed inside your network, on your hardware or your cloud account.
- Two n8n automations wired to the systems you actually use — ERP, CRM, warehouse, Slack, WhatsApp Business.
- A retrieval index over your own documents, with the ingestion script in the repo.
- Team onboarding session and a maintenance runbook: restart, update, restore, rotate keys.
- Written data-flow note for your RGPD register — what leaves the network, and proof that nothing does.
Architecture Audit
Find what breaks before it breaks.
- Timeline
- 5 business days
- Format
- Fixed scope
- Best for
- Pre-scale or post-incident
- Output
- Written report
What you get
- Two days of deep review: reliability, cost, security, and single points of failure.
- A prioritised fix list ranked by effort against impact — not a list of good intentions.
- Real postmortems from comparable systems, so the risk is concrete rather than theoretical.
- A 60-minute readout with your team, with the arguments in writing afterwards.
- Everything actionable within one sprint; larger work is quoted separately, or not at all.
Fractional SRE
Reliability without a full-time hire.
- Timeline
- Monthly
- Format
- Retainer
- Best for
- Teams of 3–30
- Notice
- One month
What you get
- Weekly review of alerts, error budget and cost — async, with a written summary.
- Incident response during business hours, with a postmortem after every outage.
- Monitoring and alert rules that fire for real problems, and stay quiet otherwise.
- Monthly reliability report: uptime, latency, on-call load, cloud spend trend.
- Architecture decisions documented as ADRs in your repo, not in my head.
Web Platforms
Real-time products, not brochure sites.
- Timeline
- 3–6 weeks
- Format
- Fixed scope
- Best for
- Internal or client-facing tools
- Handover
- Repo + docs
What you get
- A product with state: live dashboards, booking flows, streaming interfaces, operator consoles.
- React with Vite, Go or Python on the server — chosen for your team, not for my CV.
- WebSocket or SSE where the data is live, polling only where it is not.
- Core Web Vitals in the green on a mid-range phone, measured before delivery.
- Tests on the business rules, a deploy script, and a README another engineer can follow.
Automation & Integrations
Kill the manual steps between your tools.
- Timeline
- 2–4 weeks
- Format
- Fixed scope
- Best for
- Ops-heavy teams
- Delivers
- Flows + alerts
What you get
- The twenty minutes a day your team loses between ERP, CRM, warehouse and inbox — removed.
- n8n, MCP or direct API integrations, depending on what your stack can actually support.
- WhatsApp Business, Slack and email notifications where a human needs to act.
- Retries, structured logs and a dead-letter path, so a failed flow is visible instead of silent.
- A diagram of the flows and the credentials register, handed over with the code.
Self-hosting & RGPD
Your data on your hardware, with the paperwork.
- Timeline
- 2–3 weeks
- Format
- Fixed scope
- Best for
- Regulated or privacy-first
- Proof
- Restore drill
What you get
- Migration of AI and SaaS workloads back to your own infrastructure where the data file requires it.
- Proxmox, Docker or Nomad, with the choice justified in writing rather than by fashion.
- Encrypted backups and a restore drill — the restore is tested, not assumed.
- Access control, key rotation and an inventory of what runs where.
- RGPD register entries and a processing map your DPO can actually use.
Inference Optimisation
More throughput from the GPU you have already paid for.
- Timeline
- 1–2 weeks
- Format
- Fixed scope
- Best for
- Teams already serving models
- Proof
- Before / after table
What you get
- A profile of one real request path — prefill, decode, queueing, dequantisation — with the time split shown per stage.
- Serving configuration compared on your own hardware: batching, context length, parallelism, KV budget. One change at a time, measured each time.
- A throughput table at concurrency 1, 8 and 32: tokens per second, TTFT p95, and cost per million output tokens in your own setup.
- The SLO edge written down — the concurrency where p95 latency doubles without throughput gain, which is where new hardware becomes the honest answer.
- A report that separates configuration wins from kernel work, and says plainly when a hand-written kernel is not worth the effort.
What actually runs
inside each one.
Not everything I have ever touched — the pieces each engagement is built from, with the level every piece holds on the full map. If it says Production, a customer's revenue depends on it working.
- Ollama + OpenWebUI
Production - llama.cpp / GGUF
Production - Docker Compose
Production - FastAPI
Production - Qdrant retrieval
Working - RGPD data-flow note
Working
- Kafka, PostgreSQL, ClickHouse
Production - Nginx and TLS
Production - Prometheus / Grafana
Working - Nomad / Consul
Working - Ansible
Working - Kubernetes / k3s
Exposed
- Linux, systemd
Production - Docker, Nginx, Caddy
Production - Tailscale, Cloudflare
Production - Beszel, backups, restore drills
Production - Prometheus / Grafana
Working - Bash runbooks
Production
- React 18 + Vite
Production - Fastify or FastAPI
Production - PostgreSQL, SQLite
Production - WebSocket / SSE
Production - WCAG AA, keyboard focus
Production - Core Web Vitals
Working
Levels are defined once, on /stack — three of them, no stars and no percentages.
Three ways to buy
the same work.
Same standards, different commitment. The scope is written before anything starts, in all three.
Includes: a fixed scope, two weeks, one deliverable, one date. Best when the problem is known and the definition of done fits on a page.
Ends with: a running system, a runbook and a demo recording. If scope grows, it becomes a second sprint — quoted, not absorbed.
Includes: five business days, read-only access, a written report and a readout. Best before a fundraise, a migration, a hiring decision — or after an outage you would rather not repeat.
Ends with: a prioritised fix list ranked by effort against impact. You are not obliged to hire me to implement it.
Includes: a monthly block of engineering time, incident response in business hours, alerting and cost review. One month notice, no annual lock-in.
Ends with: a monthly reliability report, and a handover file that stays current for whoever comes next.
Three things that will
get expensive in a year.
Not a pitch for more work: the three constraints that usually arrive six to twelve months after the first system goes live.
Usage grows faster than the hardware plan. The bill arrives as a bigger GPU, unless somebody measured the serving configuration first.
Inference optimisationThe person who built it is now busy. Nobody owns the alerts, and the first real outage is also the first restore test.
Fractional SREA client or an auditor asks where the data goes, and the honest answer is a diagram that does not exist yet. It is cheaper to write it while building.
Self-hosting & RGPDHow the quote
actually works.
No public rate card is a decision, not a trick: a fixed price published without knowing your stack is either wrong or padded. Here is what replaces it.
The questions
that come up.
How much does it cost?
It depends on the scope, which is exactly why the first call and the written quote are free. A two-week AI Sprint and a five-day audit are different objects; quoting either without seeing your stack would be guesswork. You get one fixed number, valid 30 days, before any work starts — and the scope it covers, in writing, on one page.
Do you work under NDA?
Yes. Part of the enterprise work referenced on this site is under NDA and appears without the client's name. I can sign yours before the first call, or use a mutual NDA I keep ready.
Remote or on-site?
Both. Based in Paris, working across the EU. Most of the work is remote with a weekly demo; for infrastructure and hardware handover, one or two on-site days usually save a week of calls.
Who owns the code and the infrastructure?
You do, from the first commit. The repository is yours, the accounts are yours, the documentation is in the repo. I do not host client systems on my own hardware, and there is no licence fee to keep using what was built.
What happens after handover?
You get the repo, the runbook, a restore drill and a recorded walkthrough. If you want me to stay, that is the retainer format. If you want someone else to take over, the handover file is written for that — it is a feature, not a risk.
What do you not do?
Editorial content, photo shoots, ad campaigns, and monthly maintenance without a defined scope. Also: I do not take a project where the deadline is decided before the scope is.
Start with
a 30-minute call.
No slide deck, no discovery questionnaire. You describe the constraint; I tell you whether it is my kind of problem and what a written scope would contain.