Narrowly scoped agents
Orchestration with LangGraph, specialised agents (reading, extraction, classification, checking) and validation stages between AI and output.
Expertise · AI architecture
AI Architect
AI systems designed as systems: components with a clear contract, verifiable outputs, and operations that hold over time.
The problem I solve
An AI prototype takes a few days to build. A reliable AI system does not. Between the two lies everything the demo does not show: badly scanned documents, mixed languages, answers that must be justified, inference costs, access rights, edge cases the model handles confidently and wrongly.
My job as an AI architect is to put the model in the right place: where it adds value, surrounded by deterministic stages that check it, in an architecture a team can test, monitor and evolve.
What I design
Reliable document assistants
Versioned ingestion, permission-filtered vector search, sourced answers, audit log — up to GxP requirements.
Model pipelines
Computer vision, biomedical and genomic AI, lightweight inference (ONNX), from preprocessing to the user interface.
Architecture principles
- AI inside the system, not the other way round. The model is a component with an input and output contract, like any other.
- Verifiable outputs. Every stage produces an inspectable intermediate result; the final output is never a black box.
- A human in the loop where it matters. In healthcare and regulated settings, AI prepares and argues; the decision stays human.
- Constraints as specification. Inference budget, confidentiality, available hardware: they drive the choice between APIs, cloud and local models.
Example architectures
- Genomic pipeline — from a patient's genomic data to the medical report, through variant pathogenicity and cancer classification — accepted paper (SmartADN). See the diagram
- Multi-agent OCR — extracting multilingual documents (Arabic, French, English) into structured data, with deterministic checks. See the diagram
- Regulated RAG — document assistant for a GxP pharmaceutical environment. See the diagram
- Dataset Intelligence Engine — agents that analyse, structure, label and document datasets. See the diagram
Read: RAG in regulated environments: compliance is an architecture requirement.
Frequently asked questions
What does an AI architect do?
They decide how AI fits into a system: which stages to hand to a model, which to keep deterministic, how data flows, and how the whole thing is verified, traced and evolved. The work spans from choosing approaches (agents, RAG, specialised models) to operational constraints.
When should you use a multi-agent system rather than a single model?
When the task splits into distinct sub-problems (read, classify, extract, check) that benefit from being tested and fixed separately. Narrowly scoped agents, orchestrated for instance with LangGraph, are easier to make reliable than one generalist model asked to do everything.
How do you make RAG acceptable in a regulated environment?
By treating compliance as an architecture requirement: a versioned, controlled corpus, permission-filtered retrieval, systematically sourced answers and an audit log of every exchange. The language model is just one of the stages.
Do you work with open-source models or APIs?
Both, depending on cost, confidentiality and infrastructure constraints. I have designed fully local pipelines (Ollama, open-source models) as well as systems relying on APIs or the cloud (GCP Vertex AI).
A transformation project, an architecture to design?
Companies, startups, institutions: describe your context in a few lines. I answer personally, with a first opinion on feasibility and approach.
Propose an engagement LinkedIn ↗ GitHub ↗
Engagements in Algeria, France, Europe, North Africa and remote · Arabic, French, English · contact@ouahabi-benhenni.com