Sébastien Retoux

Engineering Manager — Backend, Data & Agentic AI Systems

About

I specialize in building robust backend APIs that scale gracefully and remain maintainable over time.
My experience covers the full lifecycle of backend systems: from designing clean service boundaries and resilient contracts, to implementing secure, high-performance endpoints that serve millions of requests reliably.

Beyond APIs, I work extensively with data pipelines, particularly using Apache Airflow to orchestrate complex workflows and ensure data quality, lineage, and timely delivery.

I have hands-on knowledge of a broad set of AWS services — EC2, ECS, CloudFront, S3, RDS, Lambda, among others — and know how to combine them into architectures that balance cost, resilience, and speed of delivery.

My current focus is agentic AI: systems where a model does not merely return text, but plans, calls tools and acts on real infrastructure. The interesting part is rarely the prompt — it is the surface you expose to the agent. I design that surface as a proper API: typed and idempotent operations, least-privilege scopes, explicit contracts, and a clear separation between what an agent may read and what it may change.

This sits much closer to my backend work than it first appears: an agent is a distributed system with a non-deterministic caller, and it needs the same discipline. Timeouts and retry budgets, structured tracing of every tool call and its cost, retrieval treated as a versioned artefact rather than a config value, evaluation suites to catch regressions a unit test never will — and a human checkpoint in front of anything irreversible. I work with tool use and function calling, MCP servers to expose internal systems to agents under controlled permissions, and RAG pipelines that reuse the data foundations described above.

I stay deliberately pragmatic here. Not every problem deserves an agent: a deterministic pipeline is cheaper, faster and easier to reason about, and I would rather ship that when it fits. Where an agent does earn its place, I care as much about the guardrails, the cost and latency envelope, and the observability as about the choice of model.

As an engineering manager, I guide teams in making pragmatic architectural choices, fostering best practices, and building a culture of learning and ownership. I enjoy helping engineers grow, while keeping delivery aligned with business goals. With AI in the loop, part of that work is now helping teams tell genuine leverage from hype, and keeping review standards intact when code is generated faster than it can be read.

Security is a constant focus in my work: from applying least-privilege principles and hardening infrastructure, to integrating authentication and authorization patterns directly into system design. Agentic systems make this sharper still — prompt injection, tool permissions and data exfiltration are threat models I design against, not afterthoughts.