Ouahabi Benhenni

Healthcare · Research

From patient DNA to medical report

Research — genomics & precision oncology · 2026

An end-to-end AI pipeline for precision medicine: from a patient's genomic data to a report a clinician can act on.

1. Genomic datapatient sequencing
2. Variant analysisannotation
3. PathogenicityNucleotide Transformer v2
4. Cancer typeclassification
5. Decision supporttherapeutic recommendation
6. Medical reportLLM-generated

Key decisions

• Every stage produces an inspectable intermediate output: the final report is never a black box.
• The pathogenicity model is progressively fine-tuned (published method, see Research).
• The LLM writes, it does not decide: it formats results from earlier stages, under human validation.

• Python
• PyTorch
• Hugging Face
• LangGraph
• LLM
• Bioinformatics

Documents · Multi-agent

Multi-agent OCR — multilingual document extraction

AI consulting · 2025–2026

A multi-agent system that turns heterogeneous, multilingual documents (Arabic, French, English) into structured, usable data.

1. Ingestionscans, PDF, images
2. Readingmultilingual OCR
3. Specialised agentsextraction per type
4. Checksvalidation & consistency
5. Structured dataAPI / database

Key decisions

• Several narrowly scoped agents rather than one generalist model: easier to test, fix and evolve.
• A deterministic check between the AI and the output: delivered data is verified, not just generated.

• LangGraph
• LLM
• Vision
• Python
• FastAPI

Pharma · Compliance

Pharma Copilot & regulated RAG (GxP)

AI consulting · strategy, roadmap and scoping

An AI strategy covering R&D, quality, regulatory affairs and data for a pharmaceutical company, and the scoping of a document assistant fit for a GxP environment.

1. Controlled corpusversioned documents
2. Indexingchunking & embeddings
3. Retrievalpermission-filtered
4. AnswerLLM with sources
5. Traceabilityaudit log

Key decisions

• In a regulated environment, compliance is part of the architecture: source versions, citations and a log of every answer.
• One roadmap per domain (R&D, quality, regulatory) rather than a single "big AI project".

• RAG
• LLM
• Vector search
• Data governance

Data · Multi-agent

Dataset Intelligence Engine

AI consulting · 2025–2026

A multi-agent system that automatically analyses, structures, labels and documents datasets — so a dataset enters a project with its ID card.

1. Raw dataset
2. Analysisprofiling & quality
3. Structuringschema
4. Labelling
5. Documentationdataset card

• LangGraph
• AI agents
• Python

Real-time · Performance

Multiplayer game with voice chat

Personal project · 2025

A multi-room architecture with real-time voice and dynamic voice permissions. It handled 218,000 requests in 3 hours on a 2 vCPU / 2 GB RAM VPS.

1. Clientsbrowser
2. SignallingSocket.io
3. Rooms & rolesvoice RBAC
4. Media SFUWebRTC / mediasoup
5. Voice streamsper room

Key decisions

• Separate signalling from media transport: each layer is sized for what it does.
• Hardware limits as specification: every byte and every connection counts on 2 GB of RAM.

• Node.js
• Socket.io
• WebRTC
• mediasoup
• Nginx
• PM2

Backend · SaaS

Multi-tenant B2B SaaS platform

SaaS company · 2025

Contributed to the architecture, API and feature design of a multi-tenant B2B SaaS platform, within an Agile team of five developers.

1. REST APIDRF · JWT
2. Tenant isolationRBAC
3. Real-timeDjango Channels
4. Async jobsCelery
5. Dashboardsanalytics

• Django 5
• DRF
• Channels
• Celery
• PostgreSQL
• JWT

EdTech · Microservices

Adaptive learning & AI-powered LMS

AI consulting · 2025–2026

A feasibility audit, then the design of a personalised recommendation and learning-tracking pipeline integrated into an existing ecosystem; then the full architecture of an AI-powered LMS (ML/LLM pipelines, APIs, data model).

1. Learning traces
2. Python microservicesREST
3. ModelsONNX
4. Recommendationspersonalised
5. Existing platformFlask / Next.js

Key decisions

• Integrate with what exists rather than replace it: AI arrives as a service, not a rewrite.
• ONNX for lightweight inference, deployed on a self-managed VPS.

• Python
• ONNX
• REST
• Flask
• Next.js
• VPS

Automation · Open source

From technical plan to catalogue

2026 · 100% open source

A pipeline that reads a technical PDF plan and produces, with no proprietary dependency, the whole chain that follows.

1. PDF plan
2. CAD drawingsgenerated
3. Suppliersidentified
4. Total cost (TCO)
5. Business plan
6. Cataloguemulti-format

• Python
• Open-source LLMs
• Ollama

Also

Other systems

• Commercial Intelligence & AI Sourcing — decision-support platform linking clients, products, suppliers, opportunities and sales.
• AI Industrial Safety — computer-vision-based industrial safety module.
• Public administration — mail management system (Express API, React, RBAC) and corruption risk mapping software.
• Streaming platform — adaptive video encoding with FFmpeg, 1,500 concurrent users.
• Multi-vendor marketplace — Turborepo monorepo, Next.js, Node.js, shared packages.
• Fully local RAG — PDF/DOCX/Markdown ingestion, vector search, inference through Ollama, no external API.

1. The problem before the model

   I start from the business, the users and what has to change for them. The model choice comes late; sometimes the right answer is not AI.

2. Constraints are the specification

   Budget, compliance, 2 vCPUs and 2 GB of RAM: these are not obstacles, they are the inputs of the architecture.

3. AI inside the system, not the other way round

   Narrowly scoped agents, verifiable outputs, deterministic stages around them. The model is a component, with a contract like any other.

4. A human in the loop where it matters

   In healthcare or regulated settings, AI prepares and argues; the decision stays human, and every answer is traceable.

5. The simplest thing that holds in production

   An architecture is judged by what it survives in real life, not on a diagram. I prefer a sober system people understand to a brilliant one they endure.

6. Make decisions explicit

   Every architecture choice is written down with its reason. That is what lets a team defend it, change it, or carry it on without me.

7. Translate both ways

   The same system must be explainable to a client, a junior developer and a steering committee. If I cannot explain it simply, it is probably split wrong.

8. Teaching is part of delivery

   120+ people mentored in bootcamps, teams of interns and juniors: a system is truly delivered when the team maintaining it understands it.

1. Understandneed, usage, constraints
2. Scopeperimeter, success criteria
3. Designarchitecture, AI choices
4. Buildteam, short iterations
5. Deployproduction, measurement
6. Hand overdocumentation, mentoring

"A successful AI project doesn't show at the demo. It shows six months later, when it is still running and the team knows why."

• An End-to-End Approach for SNV Pathogenicity Prediction via Progressive Fine-Tuning of Nucleotide Transformer-v2
  Benhenni Ouahabi (first author), Djamel Gaceb, Fayçal Touazi, Wissal Kebour, Rayane Chakib Idris, Belkacem Rouibi, Anis Ismail. Published and presented at a national conference.
• Melanoma classification with deep learning (CNNs & Vision Transformers) — paper under peer review.
• Medical AI in a hospital setting — skin cancer classification (CNN + ViT, 93% accuracy), cell detection and counting (YOLOv8), malaria detection; full pipeline from preprocessing to web interface.
• 1st place — National AI Competition (2025) — medical computer vision project using deep learning.

1. Aug 2025 — now

   AI Consultant · AI Project & Product Lead · Systems Architect

   AI consulting & engineering firm

   About ten AI projects and initiatives led from needs analysis to production, across pharma, industry, documentation, sales intelligence and education. Interface between clients, business and technical teams; coordinating teams of up to 15 people; contributed to closing major projects.

2. 2026

   AI Research Engineer — Genomic Medicine & Precision Oncology

   Biomedical research institute

   AI pipeline from genomic analysis to automated report generation; variant classification and decision support in a multidisciplinary setting.

3. 2025 — 2026

   Software Engineer (contract)

   Public administration

   Mail and correspondence management system; corruption risk mapping software.

4. Jun — Sep 2025

   Full-Stack Developer

   B2B SaaS company

   Multi-tenant B2B SaaS platform: architecture, API, feature design.

5. 2023 — 2025

   AI & Full-Stack Developer (freelance)

   Independent

   End-to-end web, desktop and AI solutions, including medical image segmentation pipelines (U-Net, YOLOv8, LinkNet, PSPNet).

6. 2022

   Co-founder — Résumily

   Academic summarisation platform

   10,000+ monthly users, official partnership with the computer science department, team grown from 13 to 43 members in two months.

7. 2018 — 2021

   Higher Technician — Database Administration

   Higher technical training

   Five management information systems delivered (stock, HR, accounting, library, internships).

1. AI diagnostic

   Feasibility audit: where AI creates value in your organisation, with which data, at what cost and with which risks.

2. Scoping & roadmap

   Requirements, use-case prioritisation, AI transformation roadmap per domain.

3. Architecture design

   AI architecture (agents, RAG, ML pipelines) and backend architecture (APIs, microservices, real-time, multi-tenant), with documented decisions.

4. Technical lead

   Leading a team from design to production: breakdown, tracking, review, deployment.

5. Team training

   Upskilling your teams on applied AI, agents and backend architecture.

A first conversation clarifies the context, the constraints and the most useful format. Write to me.