Roadmap
What to build, what to buy, what to defer. A sequence the CEO and CTO can defend, not a backlog of demos.
Fractional Head of AI · Building AlexandrIA
François Paupier
I own the AI vision for post-seed product companies, and I make it real in production.
How a mandate works · typically 1-2 days/week · capacity intentionally limited
- 1-2 days/week Typical cadence
- Monthly retainer Ongoing ownership
- Limited capacity A very small number of mandates
- 30-min diagnostic Free, no commitment
Who this is for
A fit if
- You are the CEO or CTO of a post-seed product company. You already have a product, not a pitch.
- AI is stuck on the CTO’s plate, competing with everything else that has to ship.
- You need someone who can own the vision and make production choices: models, serving, evals, data, hosting.
- Constraints are real: EU-sovereign cloud, on-prem, or air-gapped. Or you simply need RAG, agents, and observability that hold up in production.
Not a fit if
- You want a slide deck, a workshop, or “AI strategy” with no delivery.
- You are an SME in distribution or logistics, not a post-seed product company.
- You are still at idea stage, with no product and no users.
- You need a full-time hire tomorrow, or a team to staff a research lab.
What a mandate looks like
You get a counterpart, not a vendor. Roadmap, stack, production, and an exec cadence. Typically one to two days a week, on a monthly retainer. Pricing on demand, set on the diagnostic.
Stack
Models, serving, evals, and observability. LiteLLM, vLLM, Langfuse, OpenTelemetry, chosen for production, not a blog post.
Delivery
RAG, multi-agent orchestration, LLM-as-judge evals. On-prem and air-gapped copilots when the constraint is real.
Cadence
A senior seat next to the CEO/CTO. Decisions get made. Systems ship. You are not left with a PDF.
Proof
- AlexandrIA, in production. Founder & CTO of AlexandrIA, a research co-pilot with visual knowledge maps, billing, EU-sovereign hosting on OVHcloud and Scaleway, and recurring paying users. Built end-to-end: multi-agent orchestration, RAG, LLM-as-judge evals, Langfuse.
- Production systems for clients. AI lead on on-prem and air-gapped copilots: vLLM, LiteLLM, Keycloak, knowledge graphs.
- Teaching. Guest Lecturer, LLM Pretraining & Post-training, SCAI (Sorbonne Université) and Centrale Nantes. Certified Kubernetes Administrator (CKA).
- Selected talks. TensorFlow World 2019, AGIR 2024, Gendarmerie panel, Qonto HQ 2025, LLM sovereignty course, Centrale Nantes 2025.
Writing, talks, and teaching
Explore my work on production AI, generative models, and the systems around them.
Building with Claude Code
Context management and feedback loops for AI-assisted development. Slides and takeaways from Claude Builders Night, Paris, April 2026.
Generative AI architectures
A technical guide to variational autoencoders, diffusion models, GANs, and transformers, with diagrams and references.
LLM sovereignty and supply chains
A Centrale Nantes course on hardware, datasets, models, and the choices behind sovereign AI.
Build vs buy for production GenAI
A talk at Qonto HQ on integrating generative AI into products and choosing what to build or buy.
Current stack
What I work in: LLM, RAG, agents, vLLM, LiteLLM, EU-sovereign.
- TypeScript
- Next.js
- Python
- FastAPI
- LiteLLM
- vLLM
- Langfuse
- OpenTelemetry
- Postgres
- Supabase
- Docker
- Kubernetes
- Ansible
- GitHub Actions
- OVHcloud
- Scaleway
- Cloudflare
Book a 30-minute diagnostic
Free, no commitment. We check fit. I take a very small number of mandates. Pricing on demand, set on the diagnostic.
About
I am a founder in Paris. I graduated from École Centrale de Nantes with an MSc in Signal Processing & Statistics (top 10%). I build production AI systems, and I teach the parts that are easy to get wrong.
Day to day I am Founder & CTO of AlexandrIA. On the side I take a very small number of Fractional Head of AI mandates. I also lead AI on client production systems when the constraint is on-prem or air-gapped.
I lecture on LLM pretraining and post-training at SCAI (Sorbonne Université) and Centrale Nantes. From 2020 to 2024 I produced and hosted Post Mortem (24 episodes, currently on hold).