Mohamed Mahdi
ICT Engineering Student //
Solo
About
I'm an independent researcher and engineering student passionate about building intelligent systems that solve my daily problems. I love competitive programming for the mental challenge and deep learning because it tackles the toughest issues I face. When I'm not coding, you'll find me playing guitar or experimenting with new recipes in the kitchen.
If you’ve got any questions or just feel like chatting, send me an (Email) I’d love to hear from you.
Publications
March 2026
Epileptic Seizure Prediction Using Patient-Adaptive Transformer Networks
This paper introduces a novel transformer-based approach for predicting epileptic seizures from EEG signals using patient-adaptive learning. By combining self-supervised pretraining with personalized fine-tuning, the proposed model learns both general brain activity patterns and patient-specific seizure characteristics. The framework achieves strong results on the TUH EEG dataset, demonstrating high accuracy and robust F1-scores in seizure forecasting tasks. This work highlights the growing potential of transformer architectures and personalized artificial intelligence in advancing intelligent healthcare systems and improving early seizure detection.
November 2025
How Well Do LLMs Understand Tunisian Arabic?
This paper explores how well modern large language models understand Tunisian Arabic, a low-resource dialect that is often overlooked in mainstream AI systems. The authors evaluate models on tasks such as translation, transliteration, and sentiment analysis using a newly built multilingual dataset that includes Tunisian dialect, standard Arabic, and English. Their findings reveal that while current models perform fairly well in structured settings, they still struggle with informal expressions and cultural nuances specific to Tunisian speech. The study highlights a clear gap in language coverage and emphasizes the need to better support underrepresented dialects to make AI systems more inclusive and effective for real-world users.
Projects
See AllTunsiGPT: Transformer LLM Trained from Scratch
Trained a transformer-based language model from the ground up on Tunisian Arabic, gaining hands-on experience in data preprocessing, tokenization, and generative model training. Built and cleaned large multilingual datasets, improving data consistency and reducing noise. Achieved measurable improvements in generation quality on low-resource dialects through systematic evaluation and architecture refinement.
Cybersecurity AI Agent
Cybersecurity platform automating vulnerability scanning, information gathering, and exploit validation. Features natural language interface with conversation memory, PostgreSQL-backed caching, and comprehensive security assessment tools including SQL injection, XSS, and CSRF detection. Orchestrates intelligent workflows for authorized penetration testing with automatic exploit validation and severity filtering.
MahdyStudies Platform
Full-stack educational platform with dynamic resource management, course scheduling, and real-time notifications. Deployed production system serving multiple users.
Work Experience
Researcher at Toulouse Research Institute in Information Technology (IRIT)
- Worked on AI4SE (AI for Software Engineering) multi-agent systems
- Designed and proposed new architectures for multi-agent system coordination
- Conducted critical analysis of existing multi-agent approaches in the literature
- Evaluated proposed solutions using formal system properties and correctness criteria
- Contributed to research-oriented work aiming to advance multi-agent system design
R&D Engineer at EchoParrot (ONRTECH)
- Built real-time speech enhancement modules improving transcription quality and downstream reasoning accuracy for conversational AI systems
- Optimized processing components reducing latency while increasing reliability in real-time interaction scenarios with minimal supervision
- Collaborated cross-functionally with NLP team to match model requirements, sharing progress and discussing technical trade-offs
- Contributed to cloud deployment and performance tuning, ensuring high availability of continuous audio analytics
- Took ownership of debugging and performance analysis in fast-paced R&D environment
AI Agent Developer at Anqua Edge Innovation
- Developed LLM-powered natural-language analytics assistant transforming user queries into structured backend operations, laying groundwork for text-to-SQL and retail decision support tools
- Built efficient FastAPI services and SQLAlchemy-driven data pipelines enabling real-time query execution and context-aware insights
- Implemented user-behavior tracking system capturing interaction events to understand purchase patterns and enhance recommendations
- Designed clean, reliable data models supporting traceability and consistent feature extraction for what-if analysis and multi-layer assistant workflows
- Achieved functional deployment in 13 hours demonstrating rapid prototyping and efficient problem-solving
Researcher at Digital Research Center of Sfax
- Researched generative diffusion models for historical manuscript retrieval, evaluating state-of-the-art models for style transfer and handwriting generation
- Designed novel similarity measurement pipelines combining CLIP embeddings, One-DM style embeddings, and custom hybrid ViT-ResNet encoder
- Fine-tuned generative models on George Washington dataset achieving measurable improvements in historical handwriting adaptation
- Developed two-stage retrieval system (CLIP coarse retrieval + One-DM re-ranking) improving retrieval accuracy
- Conducted systematic OCR experiments with TrOCR and quantitative analysis using mAP metrics
- Contributed to research methodology and experimental documentation for scientific publication
Machine Learning Engineer Intern at Welyne
- Designed and deployed RESTful API using FastAPI for real-time human body measurement predictions with production-ready endpoints
- Collected, labeled, and preprocessed custom dataset implementing noise handling techniques for data quality assurance
- Built and evaluated ML models with Scikit-learn comparing regression and classification algorithms for optimal performance
- Applied feature engineering techniques including selection, scaling, and transformation to improve model accuracy
- Deployed models with joblib enabling real-time predictions via HTTP endpoints for business integration
- Collaborated cross-functionally to deliver API meeting technical requirements and business constraints
Extracurriculars
Organization Team - Complex Systems Engineering and AI
Conference- Contributed to the organization of a professional conference during my internship, supporting coordination and event preparation.
- Assisted with logistics and catering arrangements to ensure smooth execution of the event.
- Actively engaged in networking with researchers and professionals in my field, while presenting and discussing my work to expand visibility and collaborations.
Panel Speaker – JCI Ksar Said: Future of AI
Talk- Participated in panel discussion on the future of AI and AI agents with industry professionals
- Addressed real challenges in AI research: data ethics, identity and consent of individuals in training datasets
- Discussed technical challenges including hallucinations, trust issues, and memory limitations in AI systems
- Shared insights on moving forward and opportunities ahead for those leading in R&D
- Engaged audience on ethical considerations and practical implementation of AI technologies
Entrepreneur & Team Lead – Oreggami (INJAZ Tunisia)
Entrepreneurship- Elected by peers to lead student startup Oreggami as part of INJAZ Tunisia entrepreneurship program
- Guided team to first place in local startup competition, outperforming 7 other business ideas
- Advanced to national finals competing with 13 teams and won 'Most Innovative Idea' award
- Developed business model, coordinated project tasks, and facilitated team decision-making
- Honed leadership, public speaking, and business planning skills in real-world startup simulation
Member – G2FOSS ENIT
Club Activity- Actively participated in G²FOSS activities promoting open-source development and knowledge sharing
- Contributed to sponsorship team during League of Coders event, securing logistical support
- Supported logistics team ensuring smooth planning and execution of event operations
- Organized weekly competitive programming workshops building problem-solving skills
- Organized and created problem sets for contests at the national level
- Collaborated with students to promote open-source values and community building
Marketing Manager – Google Developer Student Club ENIT
Club Activity- Managed marketing and promotion efforts for events and workshops organized by GDSC ENIT
- Created digital content to increase visibility and engagement across platforms
- Worked with technical and creative teams to design campaigns supporting student-led tech projects
- Participated in club meetings, project planning, and outreach to the ENIT community
- Contributed to culture of learning, technology, and community impact
IEEE Mentor – Workshops & Events
Workshop- Organized breast cancer awareness web development workshop combining social impact with technical education
- Led code camp where team developed donation website for real-world community benefit
- Organized competitive programming workshops fostering algorithmic thinking and problem-solving skills
- Mentored students through hands-on projects bridging theory and practical application
