AI platform architect · Europe / Remote
I design and build private AI platforms, tune LLM inference, and ship the applications that run on top—without handing your roadmap to an external API.
Design or evolve a self-hosted control plane for language, vision, speech, embeddings and image generation—with sensible gateways, observability and capacity planning.
Benchmark the workload that matters, find the bottleneck, and choose a serving layout from evidence—not vendor slides.
Build resilient application layers for media and knowledge workflows, including long-running jobs, RAG, speech and document understanding.
Turn a fuzzy AI initiative into an actionable architecture, benchmark plan and delivery path your team can own.
02 Selected systems
$ benchmark qwen3.6-fp8 --concurrency 754,970output tok/s12.6 ms inter-token · 1.0 s p95 TTFT · 0 errors
Built and run an on-premises serving platform for language, vision, image, embedding and speech models. Fifteen teams access governed services through one gateway, with measured fallbacks and shared observability.
Designed prefix-cache-aware routing and a shared client harness. Warm requests that once waited up to 90 seconds now reach first token in under one second.
Built a PostgreSQL-backed job system with capacity-aware dispatch, leases, retries, cancellation fencing, ordered callbacks and atomic artifacts for speech and OCR.
Selected details only. Client-sensitive architecture and outcomes can be discussed privately.
03 About
I’m Jean-Baptiste, an AI platform architect and inference engineer working between deep systems engineering and useful products.
At RTBF, I operate the AI platform behind editorial tools and media workflows while supporting recommendation systems used by millions. I care about performance, reliability and technological sovereignty because production makes vague claims expensive.
I also lead AI compute and inference initiatives at Amaris Consulting. My earlier path through data engineering and video editing gives me an unusual lens: infrastructure matters, but the human workflow at the end matters more.
04 Field notes
What theoretical load tests reveal—and why KV-cache locality decides whether an agent feels instant or broken.
Read field note ↗Keys and routing are the easy part. Ownership, fallbacks and observability decide whether teams trust the platform.
In the notebookWhy leases, idempotency and atomic artifacts matter when inference lasts minutes.
In the notebook05 Ways to work together
Pressure-test the choices and leave with a prioritized delivery plan.
Typically 1–2 weeksBuild a platform, unblock inference performance, or ship a production workflow.
Fractional or full-timeBenchmark serving options and turn the results into concrete changes.
Focused engagementHave a hard AI systems problem?
Tell me what you’re building, what is stuck, and what success looks like. I’ll reply with a useful next step.
jbgidrol@hotmail.fr ↗