Hello, I'm

Soulaima
Sellemi

AI engineering student, final year

Software Engineering · ENSI, University of Manouba

LLM agents RAG Deep learning MLOps

I turn promising AI research into tools people can trust: multi-agent LLM pipelines, RAG systems and deep-learning models, evaluated on real-world data.

Looking for a PFE internship End-of-studies project · 2027 · AI / LLM engineering

Get in touch

ENSI · MANOUBA2027

Soulaima
Sellemi

AI Engineering Student

LLM agents · RAG · Deep learning

Open to PFE

CONTACT2027

Let's connect

Email: soulaima.sellemi@ensi-uma.tn
LinkedIn: in/sellemi-soulaima
GitHub: soulaimasellemi
Based in: Manouba, Tunisia
Languages: Arabic · French · English

or drag it to swing

2R&D internships
at Larodec Lab

937KGerrit changes
mined & analysed

+7.4 ptsaccuracy over a
published ICSE '26 baseline

97.0%Dice score on
PH2 segmentation

Who I am

About me

I'm a final-year software engineering student at the National School of Computer Science (ENSI), University of Manouba, after two years of Mathematics & Physics preparatory classes at IPEIB.

What drives me is the gap between a promising paper and a tool someone can rely on. In my research internships I take published methods, test them on messy, real-world data, find out where they break, and build something better.

I'm comfortable across the whole path, from training and evaluating models to shipping them with LangChain / LangGraph, Docker and Jenkins CI/CD.

Currently looking for an end-of-studies (PFE) internship in AI / LLM engineering.

location Manouba, Tunisia
degree Engineering, Software Eng. (2024–2027)
languages Arabic (native) · French (C1) · English (B2)
also Treasurer, ENSI Robotics Association

Lab work

Research experience

Summer 2026Larodec Laboratory, ENSI

AI R&D Intern

Toward AI-Assisted Decomposition of Large Pull Requests into Reviewable Changes

Large pull requests are hard to review. I tested whether a published multi-agent LLM framework could split them into smaller, reviewable changes, using real developer history instead of synthetic benchmarks.

65.0%Atomizer (ICSE 2026) on real Gerrit data

72.4%my taxonomy-routed pipeline

κ = 0.896inter-rater agreement on taxonomy

Mined 937,327 Gerrit changes from 7 ecosystems into 142 validated change-split cases, each traced to the developer's own words.
Built a taxonomy of why and how developers split changes (κ = 0.896 and 0.769).
Diagnosed over-splitting as the main cost driver, then designed a taxonomy-routed pipeline that beat the baseline.

Multi-agent LLMsLangGraphMining software repositoriesEmpirical evaluation

Read the full case study

Summer 2025Larodec Laboratory, ENSI

Deep Learning Intern

AC-MambaSeg: Hybrid CNN-Mamba Model for Skin Lesion Segmentation

Reproduced and extended a hybrid architecture combining CNNs with Mamba state-space blocks, CBAM and attention gates for medical image segmentation, in PyTorch Lightning.

97.0%Dice on PH2 (CNN baseline 95.0%)

94.2%IoU on PH2 (baseline 91.0%)

90.5%Dice on ISIC 2018, ahead of SM-UNet