AI architect in Paris · Enterprise AI & cloud
I design and build the platforms behind enterprise AI: model gateways, RAG that cites its sources and agents that act within your rules. Hands-on, on AWS, for teams that have to pass security review.
One request through the platform (illustrative)
Delivered for and with
What I fix
The model is rarely what stalls an AI project. It is the things around it: access, trust, control and the cloud underneath. Those are the parts I build.
Every team wires up its own model.
A central AI gateway on Amazon Bedrock enforces EU-only inference and PII filtering, then routes each request to the right model. Products stop depending on any single one.
The assistant makes things up.
RAG on PostgreSQL with pgvector, hybrid search and reranking. Tenant isolation is enforced in the database itself.
Nobody trusts an agent to act.
One MCP action server with versioned tools, human approval on every write and a decision log you can replay. The agent never holds more rights than the user.
Nobody can tell if the last change made it better.
Golden sets and evaluation gates in the pipeline, plus tracing of every prompt and answer, so a regression fails the build instead of reaching users.
It runs on one server. Then real traffic arrives.
Multi-AZ platforms on AWS and GCP defined as code, with Kubernetes, GitOps and CI/CD that your teams can run without me.
You want to run open-source models yourself.
Distributed inference on a GPU cluster, with the GenAI pipelines, fine-tuning and automation that go around it.
Toolbox
Technologies from the platforms above and from my own projects.
Selected work
Six engagements, each told by the problem and the outcome.
Lead architect of the shared platform behind four AI product capabilities, from model access to agent actions.
40+ projects on automated CI/CD and 50% fewer manual interventions, on HDS-compliant infrastructure.
Multi-AZ AWS infrastructure defined as code, ensuring 99.9% availability for the GENWIN web platform.
Spring Boot microservices on managed GCP services, with observability built in from the start.
Vespa AI vector search cut email response time by 70% on a medical follow-up platform built with LLMs and RAG.
An interactive media generation platform running open-source models on a RunPod GPU cluster, with GenAI pipelines, agents and n8n workflows.
How it works
Four steps, each with something you can hold in your hands.
A 30-minute call. You describe what you are building and what is blocking it. I tell you honestly whether I can help.
I review what exists and design the target: a reference architecture and decision records that your team can challenge.
I work inside your team on the hard parts: the gateway, retrieval, agent actions, the pipelines. You get code, not slides.
Evaluation gates, runbooks and dashboards, so your team can run the platform and change it without me.
Field notes
Design guides with the decisions and the trade-offs, not the hype.
A design guide to an enterprise AI gateway on Amazon Bedrock: EU-only inference profiles, SCP enforcement, Guardrails PII filtering and model routing.
A reference architecture for an enterprise AI platform: AI gateway, retrieval, agent actions, evaluation and identity, and the order to build them in.
A design guide to governed agent actions over MCP: versioned tool manifests, human validation on writes, idempotency, decision logs and OAuth2 Token Exchange.
Last updated:
Tell me what you are building. After a first conversation you will know whether I can help and what I would do first.
+33 6 02 73 49 22 me@osmanrami.fr