Naoufel Azizi — AI & ML Engineer

Chapter 01 · The Finding

Most of the apparent skill was the dataset, not the disease.

I'm a final-year Master's student in Machine Learning, NLP & Data Science (MIND) at Sorbonne Université, Paris, with a background in CPGE mathematics and computer science from Algeria.

My research at LIP6 (CNRS) is on medical image analysis: classifying colorectal polyps from endoscopic images, supervised by Garance Lucas under Bertrand Granado. What I found there shapes how I work — a model that scores well on a held-out slice of one dataset may have learned the dataset rather than the disease. I care about results that survive a change of source, and about reporting per-stratum instead of per-headline.

Beyond research, I build software — from CakePHP web applications to Flutter mobile apps and containerised deployments on cloud infrastructure. I like elegant solutions to hard problems, whatever the stack.

Fig. 01 — Leave-one-dataset-out−81.0%

balanced acc. above chance — source in training +0.159

balanced acc. above chance — source held out +0.031

three-class macro F1, in-distribution 0.739

sessile recall, held-out sources 0.000–0.065

4× minority augmentation, macro F1 gain +0.0013

Fig. 01 — Mean balanced accuracy above chance across seven public colonoscopy sources, measured twice: with the source represented in training, and with it held out entirely. Most of the apparent skill is recognition of the dataset, not the pathology. LIP6 (CNRS), 2026.

Languages spoken — French, English, Arabic

93k

Images in the polyp benchmark I built at LIP6

Mar 2027

Available — 6-month research internship

Paris

Based — open to new opportunities

Chapter 02 · The Journey

Experience & Education

Jun–Jul 2026 · Paris

Research Intern (M1) — Medical Image Analysis

LIP6 / Sorbonne Université

Classifying colorectal polyps — adenoma, hyperplastic, sessile serrated — from endoscopic images, and testing whether what the models learn is the lesion or the dataset. Supervised by Garance Lucas (doctoral researcher) with Bertrand Granado as academic tutor and Hubert Naacke as referent.

PyTorch OpenCV Computer Vision Medical Imaging Python

Apr–Jul 2025 · Clamart

Software Developer

Diese Telecom

Full-stack development on "logidiese", a residential management CakePHP application. Redesigned multiple UI views with a modern dashboard aesthetic, added SAV status tracking, fixed longstanding display bugs, and containerised the entire stack on Azure with Docker Compose.

CakePHP PHP MySQL Docker Azure

2025–2027 · Paris

M2 MIND — Machine Intelligence & NLP Data

Sorbonne Université · UFR 919

Advanced coursework in deep learning, natural language processing, information retrieval, and computer vision. Final-year specialisation in medical image analysis research.

NLP Deep Learning Computer Vision HuggingFace

2022–2024 · Algeria

CPGE — Mathematics & Computer Science

Classes Préparatoires aux Grandes Écoles

Intensive preparatory program in mathematics, physics, and computer science — building the analytical foundation for advanced machine learning and AI research.

Mathematics Algorithms CS Fundamentals

Chapter 03 · The Work

Selected Projects

The evidence behind the story — research first, engineering next to it.

Graph RAG — extracted knowledge graph, cluster overview

Graph RAG — entity neighbourhood, Nicolaus de Ultricuria

Graph RAG — entity neighbourhood, Adam de Cambrai

Graph RAG — entity relations, works and travels

Graph RAG — Knowledge Graph Extraction

An LLM-based pipeline that extracts structured knowledge graphs from unstructured historical text — entities (people, works, institutions, places) and typed relations (authored, taught at, travelled to, member of) — applied to the Studium Parisiense dataset on medieval Parisian scholars. Powered by gemma-3-27b-it served with vLLM on A100 GPUs, with an entity-deduplication stage and interactive graph exploration over the extracted network.

Colonoscopy frame — sessile polyp, dataset sample 1

Colonoscopy frame — pedunculated polyp, dataset sample 2

Colonoscopy frame — flat lesion, dataset sample 3

Colonoscopy AI Research

Built a 93,156-image, lesion-annotated benchmark from the 8 of 23 public colonoscopy datasets that actually label polyp class — after removing 16,432 duplicate files (15% of everything on disk) that would have put the same physical lesion on both sides of every split. Designed lesion-disjoint splits and a leave-one-dataset-out protocol across three frozen backbones on a SLURM cluster. Three-class macro F1 reaches 0.739 in-distribution — but balanced accuracy above chance falls 81% once the test source is held out, and recall on the hardest class collapses to near zero. The models were substantially recognising which dataset an image came from, not the pathology.

Sample frames — PolypsSet / GianaChallenge research datasets

Query Variation Generators

NLP evaluation framework measuring IR system robustness under natural query reformulations. Key finding: ~−20% retrieval performance loss under variation on the ANTIQUE dataset — exposing fragility in standard information retrieval benchmarks.

LLM Inference on HPC Cluster

Set up a local LLM inference node on a university HPC cluster using vLLM 0.8.5 with Qwen2.5-32B-Instruct (GPTQ Int4). Resolved a chain of non-root environment constraints, proxy issues, and API format bridging between completion backends.

Logidiese admin dashboard — Maison JN, services overview

Logidiese — Residential Management

Modernised a legacy CakePHP residential management platform for Diese Telecom. Rebuilt the admin views with a contemporary dashboard aesthetic (residences, equipment, residents, events), added SAV status tracking, resolved longstanding display bugs, and containerised deployment on Azure.

SmartDiese app — smartphone call settings

SmartDiese — Intercom & Access App

Flutter mobile app for Diese Telecom pairing residential intercoms with residents' smartphones — remote door release over VoIP, call-forwarding configuration, and custom ringtones. Upgraded for Dart 3.8.0 and implemented native iOS incoming-call handling with PushKit / CallKit (flutter_callkit_incoming) for lock-screen answer support.

Home Tech — E-Commerce Platform

A React single-page e-commerce storefront: homepage carousel, product catalogue with filtering and ratings, a product detail view with quantity selector and thumbnail gallery, and a full auth flow (sign up, login, animated panel switch) built in the style of a real electronics retailer.

Naoufel
Azizi

Most of the apparent skill was the dataset, not the disease.

Experience & Education