Mouad El Abbassi

Leyton CognitX

AI Engineer / Data Science Intern

Jun–Sep 2026

An evidence-grounded LLM intelligence layer over transcribed business calls.

• Benchmarked long-context extraction pipelines using adaptive chunking, structured outputs, Pydantic, and Elasticsearch. Information capture improved from 67.5% to 81.7%.
• Explored entity-graph organization; the retained call-intelligence implementation uses structured Elasticsearch documents with traceable evidence.
• Built an agentic QA workflow with intent classification, typed planning and routing, DAG execution, Elasticsearch retrieval, and FastAPI.

Information capture: 67.5% → 81.7%. Equal-weight mean across two evaluated calls in the supplied internship report. This is a limited benchmark, not a general production estimate.

Skills: Pydantic, Elasticsearch, FastAPI, Adaptive chunking, Structured outputs, Knowledge Graphs, DAG execution, Python, Git, GitLab, Graph Theory, Agentic Workflows, Collaborative Problem Solving, Benchmarking, Knowledge Engineering, Domain-Specific Languages, Agentic AI Development, AI Agents, Prompt Engineering, Extract, Transform, Load, Large Language Models, Data Science, Teamwork, Problem Solving

ENSA Khouribga internship

Internship

Jun–Oct 2025

A scraping and cleaning pipeline, parameter-efficient LLM fine-tuning, and retrieval-augmented generation.

• Built a scraping and cleaning pipeline.
• Fine-tuned Llama 3.1 8B with parameter-efficient training.
• Implemented RAG with ChromaDB.

Skills: Python, BeautifulSoup, Web Scraping, Data Cleaning, Llama 3.1 8B, Parameter-efficient training, PEFT/LoRA, QLoRA, Transformers, ChromaDB, Sentence Transformers, Artificial Intelligence, Natural Language Processing, Hugging Face, Kaggle, Jupyter, Fine-tuning, Retrieval-Augmented Generation, React, Node.js, Data Visualization, Frontend Development, Machine Learning

AI-Driven Microgrid & Explainable EMS

Reinforcement learning · Sustainable energy

Anticipatory energy control for a 50-home Atlas village, combining solar forecasting, reinforcement learning, and local AI explanations.

• Won 1st Prize in the GREENTECH Innovation Competition for a hybrid microgrid design in the Khénifra region of Morocco.
• Built a digital twin coupling photovoltaic generation, lithium-ion storage, a 1,500-ton thermal sand battery, and building heat-loss physics.
• Sourced three years of hourly NASA POWER weather data. Synthesized household demand with staggered usage, seasonal variation, and Ramadan load patterns.
• Used Random Forest for solar forecasting and XGBoost for electrical and thermal demand, with cyclic time encoding and temporal splits to limit future-data leakage.
• Trained a PPO policy in a custom Gymnasium environment with ten observations, three continuous actuator commands, comfort penalties, battery-health penalties, and deterministic thermal safety limits.
• Separated the fast control loop from background explanations using FastAPI, LangChain, and Mistral 7B through Ollama. Backend status tags ground the explanations in decoded telemetry.

GREENTECH Innovation Competition: 1st Prize. Award reported in the supplied engineering write-up.

Hourly weather observations: 26,280. Three years of NASA POWER data used to build the simulation.

Competition winner · Simulation

Predict

Forecast sunlight and village demand before a cold night arrives.

Dispatch

Use lithium-ion storage for fast fluctuations and preheat sand storage with midday surplus.

Explain

Translate the policy action into an operator explanation without blocking control.

Forecasts become useful when they change a decision. This project connects predictive models to physical control and understandable operator feedback.

Skills: Python, PyTorch, Scikit-learn, Random Forest, XGBoost, Gymnasium, PPO, Stable-Baselines3, FastAPI, LangChain, Ollama, Mistral, NASA POWER API, Digital Twins, Feature Engineering, Model Evaluation, Prompt Engineering, Reinforcement Learning, Deep Learning, Deep Reinforcement Learning, Teamwork

Technical architecture & engineering report (TXT, 15 KB)

Mira · Moroccan Darija Language Model

Low-resource NLP · Language adaptation

A low-resource language adaptation pipeline for natural Moroccan Darija across Arabic script, Arabizi, and mixed French or English conversations.

• Developed the original transcript-derived Q/A pipeline, including the earlier 70,000-pair collection and Label Studio review, then extended the work into the Mira dataset and training experiments.
• Consolidated CSV sources into JSONL, removed weak or duplicated answers, bucketed by script and intent, and tracked augmentation provenance. Integrated multi-turn conversations from Darija-SFT-Mixture.
• The current report documents a 40,025-conversation chat dataset and a deduplicated raw Darija corpus of 75,421 utterances, assembled into 5,220 clean continued-pretraining chunks.
• Tested Mistral and Gemma supervised fine-tuning, including full-dataset, multi-turn, and Arabizi-focused runs. Gemma SFT is the current baseline; native Darija fluency remains an experimental limitation.
• Prepared the next experimental sequence: fresh Gemma 3 4B, continued pretraining, assistant-only SFT, then retrieval and preference tuning. RAG and DPO in the current Mira report are proposed next stages.
• The earlier portfolio brief also records LangChain RAG and MLflow work. The newer report refines the evaluation plan around native-speaker naturalness, MSA leakage, Arabizi coherence, hallucinations, and multilingual retention.

Current chat dataset: 40,025. Conversation records in dataset_reformed.jsonl, as reported in Chatbot_Report.

Clean CPT corpus: 5,220 chunks. Prepared for the next language-adaptive pretraining experiment.

Experimental · SFT baseline

Curate the language

Audit script, intent, duplicates, synthetic wrappers, and conversation structure.

Test the assumption

Compare SFT runs and diagnose why assistant style does not guarantee native Darija fluency.

Adapt before aligning

Prepare language pretraining before the next SFT, retrieval, and preference experiments.

A training loss is only one signal. Low-resource language work also needs native-speaker evaluation and careful separation of fluency from factual grounding.

Skills: Python, Pandas, Transformers, Hugging Face, PEFT/LoRA, QLoRA, TRL, Mistral, Gemma, Label Studio, LangChain, MLflow, Data Cleaning, Feature Engineering, Model Evaluation, Natural Language Processing, Large Language Models, Fine-tuning, Tokenization

Mira language-model project report (PDF, 756 KB)

MouadVision · Commerce & Predictive Analytics

Full-stack engineering · Applied ML

2025–2026

A multi-role commerce platform that connects products, orders, seller revenue, stock management, and predictive sales insights.

• Built a Spring Boot and Java backend, React and TypeScript frontend, MySQL persistence, and an independent FastAPI machine-learning service communicating through REST APIs.
• Cleaned 708 source product records to 655 usable rows, then enriched 513 products through web scraping after excluding unavailable product links.
• Implemented JWT authentication and role-based access for administrators, analysts, sellers, and buyers, with protected routes and separate dashboards.
• Developed catalogue filtering, cart and order workflows, seller product approvals, stock alerts, reviews, notifications, and automated seller/platform revenue allocation.
• Integrated an XGBoost bestseller classifier, experimental ranking prediction with synthetic labels, and category-based statistical price recommendations.
• Presented analytics with Chart.js and Recharts and added PDF, Excel, and CSV exports. Payments are simulated; semantic search is documented as future work.

Enriched product catalogue: 513 products. From 708 initial source records after cleaning and enrichment.

Distinct access roles: 4 roles. Administrator, analyst, seller, and buyer.

Academic application

Prepare the catalogue

Clean, validate, enrich, and persist product data.

Run the business

Connect four roles to orders, approvals, inventory, and revenue.

Support decisions

Surface predictions and price recommendations alongside business analytics.

Applied ML needs a complete product around it: consistent data, clear permissions, usable predictions, and dependable business workflows.

Skills: Java, Spring Boot, Spring Security, Spring Data JPA, React, TypeScript, Tailwind CSS, MySQL, JWT, Role-based Access Control, Python, FastAPI, Pandas, NumPy, Scikit-learn, XGBoost, Chart.js, Recharts, REST APIs, SQL, Web Scraping, Data Cleaning, Model Evaluation, Git, GitHub, Spring Framework, Microservices, Natural Language Processing, Machine Learning, Full-stack Development

Complete MouadVision report (LaTeX source) (TEX, 83 KB)

Evidence-Grounded Call Intelligence

Agentic QA · Information extraction

Jun–Sep 2026

A typed extraction and conversational QA system that makes long business calls searchable, traceable, and easier to evaluate.

• Co-developed the project with Achraf Zahid at Leyton CognitX, beginning with manual analysis of 30 real calls.
• Designed a typed Summary contract with timestamped evidence, deterministic metadata, and a separate retrieval-oriented semantic representation.
• Benchmarked prompts, models, and chunking strategies; retained adaptive chunking, independent extraction, and controlled merging with Pydantic validation.
• Improved mean common-information capture from 67.5% to 81.7% across the two evaluated calls. The report explicitly limits generalization from this small benchmark.
• Developed intent classification, typed planning and routing, DAG execution, Elasticsearch retrieval, conversational state, and grounded answer generation behind FastAPI endpoints.
• Explored entity-graph organization; the retained implementation stores structured call documents in Elasticsearch. GraphRAG appears as related research rather than the implemented retrieval system.

Mean capture on two evaluated calls: 67.5% → 81.7%. A 14.2 percentage-point improvement on the report’s common-information ledger; not a broad production estimate.

Internship project

An answer becomes more useful when its claims can be traced back to the source conversation and tested against a shared evaluation contract.

Skills: Python, Pydantic, Elasticsearch, FastAPI, Adaptive chunking, Structured outputs, Knowledge Graphs, DAG execution, REST APIs, Model Evaluation, Git, GitLab, Graph Theory, Agentic Workflows, Collaborative Problem Solving, Benchmarking, Knowledge Engineering, Domain-Specific Languages, Agentic AI Development, AI Agents, Prompt Engineering, Large Language Models, Data Science

Internship engineering report (PDF, 23.4 MB)

ENSAKH · Intelligent Educational Assistant

Domain adaptation · Retrieval

A university assistant combining an institution-specific Llama model with a searchable knowledge base for students and staff.

• Scraped ENSA Khouribga institutional information with Requests and BeautifulSoup, including programmes, courses, administration, and faculty information.
• Reduced approximately 12,000 scraped entries to about 2,100 cleaned examples and stored structured metadata separately from retrieval embeddings.
• Fine-tuned Llama 3.1 8B Instruct using parameter-efficient LoRA and 4-bit quantization. Used topic-aware splits and similarity filtering to reduce overlap between training and test examples.
• Built ChromaDB retrieval with sentence-transformer embeddings and an ingestion workflow for newly added institutional documents.
• Designed a FastAPI query pipeline connecting intent detection, vector search, context construction, and generation. The presentation describes both implemented components and integration work in progress.
• Designed user feedback collection and an RLHF improvement loop. The report does not provide a separate evaluated RLHF training result.

Curated institutional examples: ~2,100. Cleaned from roughly 12,000 scraped text entries.

Fine-tuning & RAG prototype

Domain fine-tuning teaches response behavior; retrieval supplies institutional information that can be refreshed independently.

Skills: Python, BeautifulSoup, Web Scraping, Data Cleaning, Llama 3.1 8B, PEFT/LoRA, QLoRA, Transformers, Hugging Face, ChromaDB, Sentence Transformers, FastAPI, Model Evaluation, Artificial Intelligence, Natural Language Processing, Kaggle, Jupyter, Fine-tuning, Retrieval-Augmented Generation, React, Node.js, Data Visualization, Frontend Development, Machine Learning

ENSAKH chatbot presentation (PDF, 6.2 MB)

Emotion & Stress Intelligence

NLP · Comparative model engineering

2025–2026

A complete text-analysis application comparing five model families across 28 emotions and stress detection.

• Worked in a team with Famakan Camara and Mohamed Agoulzi to compare logistic regression, LinearSVC, BiLSTM, BERT, and DistilBERT.
• Prepared GoEmotions and Human Stress Prediction data with cleaning, tokenization, WordNet augmentation, and train/validation/test partitions.
• Compared macro F1 for emotion classification, binary stress metrics, confusion matrices, model size, training cost, and inference latency.
• The report records BERT at 84% macro F1 for emotions and LinearSVC at 83.55%. Stress results on augmented data are reported separately from DistilBERT’s original-data experiment.
• Integrated prediction, batch inference, model comparison, history, and temporal stress analysis into a Flask API, React frontend, SQLite database, and Recharts visualizations.
• Added rule-based pattern scoring and alerts for research exploration. The application is educational and does not provide medical diagnoses.

Emotion labels: 28 classes. GoEmotions classification task.

Reported BERT emotion macro F1: 84%. Report-specific experiment on augmented data; clinical validity was not established.

Educational & research prototype

Model selection is a tradeoff between quality, latency, and resources. Comparable evaluation matters as much as the architecture.

Skills: Python, Scikit-learn, PyTorch, Transformers, BERT, DistilBERT, BiLSTM, TF-IDF, Support Vector Machines, Logistic Regression, NLTK / WordNet, Flask, React, SQLite, Recharts, REST APIs, Data Cleaning, Model Evaluation, Natural Language Processing, Machine Learning, Deep Learning, Data Science, Benchmarking

Emotion and stress project report (PDF, 5.6 MB)

Mira · Darija Conversation Interface

Angular · Conversational UX

2025–2026

An Angular conversation interface that gives the Darija model project a distinct visual identity and a usable chat experience.

• Contributed to a six-person team organized in three pairs, working with Mohamed Agoulzi on the shared Mira project.
• Built a component-based Angular frontend with TypeScript and SCSS, a welcome view, chat messages, message submission, and a conversation sidebar.
• Created responsive layouts and the Mira visual identity, then validated the interface through compilation and functional interaction checks.
• The current interface uses simulated assistant responses. Next.js backend, PostgreSQL storage, Hugging Face model integration, and authentication are documented integration plans rather than completed features.

Frontend prototype · Simulated responses

A working interface can make the intended interaction concrete while the model and service integrations evolve.

Skills: Angular, TypeScript, SCSS, HTML, CSS, Responsive Design, Component Design

Angular frontend project report (PDF, 986 KB)

MovieLens · Distributed Hadoop Analytics

Big data · Distributed processing

2025–2026

A Docker-based Hadoop cluster running Java MapReduce analyses over the MovieLens ml-32m dataset.

• Built the project with Ikram El Menhi and Abderrahim Azeroual, moving from a virtual-machine approach to a reproducible container setup.
• Configured HDFS, YARN, NameNode/DataNode services, and resource management, then validated the cluster with a WordCount job.
• Loaded MovieLens CSV files into HDFS and implemented four Java MapReduce jobs: average movie ratings, rating distribution, frequent tags, and movie popularity.
• Used combiners to reduce shuffle work and automated ingestion, compilation, job execution, and output retrieval with a shell pipeline.
• Converted the aggregated results into a Python dashboard and discussed the local Docker cluster’s limits: distributed services on one host do not establish multi-machine fault tolerance.

MovieLens dataset edition: ml-32m. Dataset-scale context; four batch analysis jobs are documented.

Containerized academic cluster

Distributed data engineering includes the environment, data movement, job execution, and readable output, as well as the computation.

Skills: Java, Hadoop, HDFS, YARN, MapReduce, Docker, Linux, Bash, Python, Pandas, Matplotlib, Distributed Systems, Data Visualization

Hadoop cluster and analysis report (PDF, 960 KB)

MultiMarket · NoSQL Data Warehouse

Data engineering · Business intelligence

2025–2026

An event-driven ELT pipeline turning simulated web, mobile, and retail orders into a unified MongoDB warehouse and business dashboards.

• Worked with Ikram El Menhi on a multichannel commerce case study using three Python generators and 1,500 simulated orders.
• Developed file-event ingestion with Watchdog and PyMongo, multithreaded processing, syntactic and semantic validation, exponential retries, and circuit-breaker handling.
• Modeled orders in MongoDB and used native aggregation for revenue by channel, top products, cancellation rates, and average revenue per order.
• Used indexes to improve query performance and documented Power BI, MongoDB Charts, and Streamlit views for different analytical audiences.
• The report records 99% transaction success and 45 ms average processing per file in the local simulation. These measurements describe the academic workload, not a production service guarantee.

Simulated multichannel workload: 1,500 orders. Generated web, mobile, and physical-store transactions.

Visualization platforms: 3. Power BI, MongoDB Charts, and Streamlit.

Simulation-based academic system

Trustworthy dashboards begin with ingestion that validates, retries, and preserves the context of every incoming record.

Skills: Python, MongoDB, PyMongo, Watchdog, Pandas, Plotly, Power BI, MongoDB Charts, Streamlit, Docker, Git, ELT Pipelines, Data Validation, Database Design & Data Modeling, Data Visualization, NoSQL, Data Engineering, Data Analysis

MultiMarket Big Data Warehouse report (PDF, 1.8 MB)

Movie Discovery · Item-to-Item Recommendations

Collaborative filtering · Interactive ML

2025–2026

A movie discovery tool that finds similar films from rating patterns, tolerates misspelled titles, and handles cold-start searches.

• Built with Mohamed Agoulzi using movies.csv and ratings.csv as explicit user-feedback data.
• Filtered sparse or noisy interactions and built a sparse movie-by-user matrix with SciPy COO/CSR representations.
• Used Scikit-learn nearest-neighbor search with cosine similarity for item-based collaborative filtering.
• Added Difflib fuzzy-title matching, duplicate and self-recommendation handling, a popular-movies fallback, and caching for the Streamlit interface.
• Evaluated recommendations qualitatively and documented sparsity, popularity bias, and cold-start limitations. Quantitative ranking metrics are future evaluation work.

Streamlit application

Practical recommendation quality includes the input experience and useful fallbacks when the model has little evidence.

Skills: Python, Pandas, NumPy, SciPy, Scikit-learn, K-Nearest Neighbors, Collaborative Filtering, Sparse Matrices, Streamlit, Data Cleaning

Movie recommendation project report (PDF, 799 KB)

Laptop Market · Data Preparation & Exploration

Exploratory analysis · Statistical computing

May 2025

A reusable Python analysis pipeline for understanding how laptop specifications, brands, and form factors relate to price.

• Worked with Chaimae El Mounjali on a dataset of 1,303 laptops and 12 variables, starting with missing values, duplicate rows, inconsistent names, and mixed units.
• Normalized RAM, storage, weights, and price fields; handled missing values; removed exact duplicates; and inspected outliers with IQR-based rules.
• Produced univariate and bivariate views including distributions, box plots, correlations, scatter plots, categorical comparisons, and Q-Q plots.
• Used SciPy for Shapiro-Wilk, chi-square, and ANOVA analyses, Statsmodels for mosaic plots, and Scikit-learn for scaling.
• Automated output folders and command-line parameters to make the analysis rerunnable. The source reports inconsistent final row counts, so the portfolio uses the verified starting dataset size.

Starting dataset: 1,303 laptops. Twelve technical and commercial variables in the supplied report.

Reproducible analysis

Cleaning units and checking distributions can be more valuable than adding a predictive model to data that is not yet understood.

Skills: Python, Pandas, NumPy, Matplotlib, Seaborn, SciPy, Statsmodels, Scikit-learn, Data Cleaning, Exploratory Data Analysis, Statistical Analysis & Testing, Data Visualization, Feature Engineering, Data Analysis, Data Science

Laptop analysis report (LaTeX source) (TEX, 35 KB)

Grand Heaven · Hotel Booking Platform

Flask · Full-stack web development

Nov–Dec 2024

A responsive hotel booking platform where guests can search for rooms, check availability, reserve a stay, and manage bookings.

• Designed and developed the dynamic frontend with HTML, CSS, and JavaScript, focusing on clear navigation and a responsive reservation experience.
• Built Flask server-side operations and SQL persistence for room records, reservations, and user accounts.
• Implemented user authentication, real-time availability checking, and booking management through an intuitive interface.
• Developed room-management features to support the hotel inventory and reservation workflow.

Web application

A booking flow brings interface design, authentication, relational data, and availability logic together in one end-to-end product.

Skills: Python, Flask, HTML, CSS, JavaScript, SQL, Responsive Design, User Authentication, Database Design & Data Modeling, Full-stack Development, Web Development, REST APIs, User Interface Design

Sales Analysis · Desktop Dashboard

JavaFX · Data visualization

Apr–May 2025

An interactive JavaFX application that turns database sales records into charts, performance metrics, and business insights.

• Created a desktop application in JavaFX with dynamic charts and graphs for exploring sales trends and performance.
• Implemented JDBC database connectivity to retrieve and process sales data for the interface.
• Designed real-time data updates, intuitive navigation, and clear data presentation to support business decisions.

Desktop application

Good analytical software makes the path from a database record to a business insight easy to follow.

Skills: Java, JavaFX, JDBC, SQL, Data Visualization, Dashboard Design, Component Design, Desktop Application Development, User Interface Design, Data Analysis

WebHost Pro · Linux Web Infrastructure

Linux administration · Web hosting

A web hosting lab connecting Apache configuration, a responsive PHP website, network services, security, and operational monitoring.

• Worked with Abderrahim Azeroual, Khalid Abouelfaraj, and Mohamed Agoulzi on a Linux web-server installation and configuration project.
• Configured Apache, PHP, virtual hosts, local domain resolution, and a WebHost Pro frontend using HTML, CSS, JavaScript, and Bootstrap.
• Developed dynamic PHP pages and a contact-form handler, with responsive navigation and semantic page structure.
• Configured DHCP services and leases, UFW firewall rules, Fail2Ban, permissions, service supervision, and diagnostic commands.
• Documented monitoring, performance checks, backups, restoration, and troubleshooting. The report discusses Apache/Nginx hosting; its detailed implementation centers on Apache.

Academic infrastructure lab

Deploying a website includes the operating system, name resolution, access controls, monitoring, and a recovery path.

Skills: Linux, Apache, PHP, HTML, CSS, JavaScript, Bootstrap, Bash, Network Administration / DHCP, Server Security, Responsive Design, Web Development, File Management

Web server installation and configuration report (PDF, 656 KB)

LLMs as Algorithm Designers

Research · Operations research × Generative AI

Investigating LLM-augmented evolutionary algorithms for automated metaheuristic design and university course timetabling.

• Research title: Automated Metaheuristic Design via LLM-Augmented Evolutionary Algorithms: Theory, Hybridization, and Application to NP-Hard Constraint Satisfaction Problems.
• Studied linear programming, feasible regions, objective functions, discrete complexity, and the limits of exhaustive search in combinatorial scheduling.
• Formulated university timetabling with mandatory hard constraints and weighted soft preferences, using ITC and PATAT benchmarks as evaluation references.
• Designed hybrid search combining genetic algorithms for global exploration with simulated annealing or iterated local search for local improvement.
• Reviewed LLaMEA and EVOLER architectures, where an LLM generates or modifies executable heuristic code and receives fitness, runtime, convergence, and error feedback.
• Analyzed sandboxed evaluation, AST validation, execution timeouts, and traceback-based repair as safeguards for generated algorithms.
• The supplied work documents research synthesis, mathematical modeling, and an evaluation architecture. It does not establish a published paper or benchmark superiority.

Literature synthesis & framework design

Model the constraints

Represent scheduling rules as feasibility checks and weighted penalties.

Evolve the algorithm

Ask an LLM to vary search operators, crossover logic, and cooling schedules.

Close the feedback loop

Evaluate candidate code in a sandbox and return scores and errors to the next generation.

The research question shifts from finding a good schedule to discovering a better algorithm for finding schedules.

Skills: Python, Operations Research, Linear Programming, Constraint Modeling, Genetic Algorithms, Simulated Annealing, Iterated Local Search, LLM-Augmented Algorithm Design, Model Evaluation

Research & engineering write-up (TXT, 14 KB)

AI Fundamentals

DataCamp

AI Fundamentals credential from DataCamp.

AI Engineer for Developers

DataCamp

Mar 2026

AI Engineer for Developers credential from DataCamp.

Data Engineer

DataCamp

Mar 2026

Data Engineer credential from DataCamp.

IBM AI Engineering Professional Certificate

IBM

Jul 2026

AI Engineering Professional Certificate from IBM.

Deep Learning with PyTorch, Keras and TensorFlow

IBM

Deep Learning with PyTorch, Keras and TensorFlow credential from IBM.

Machine Learning with Python

IBM

Machine Learning with Python credential from IBM.

CODEX ENSAKH Club

Volunteering · Organisateur de Forum

Sep–Nov 2024

Contributed to organizing a forum with CODEX ENSAKH Club, bringing communication, teamwork, and problem solving into a campus event.

• Volunteered as Organisateur de Forum from September to November 2024.
• Recognition certificate shown in the supplied volunteering record.

Skills: Communication, Problem Solving, Teamwork, Event Organization

DATAVERSE

Volunteering · Logistics Assistant

Sep–Dec 2024

Supported DATAVERSE as a Logistics Assistant, contributing to event logistics and team coordination.

• Volunteered as Logistics Assistant from September to December 2024.
• Recognition certificate shown in the supplied volunteering record.

Skills: Event Logistics, Teamwork

McDonald’s

Équipier Polyvalent · Customer service

Developed customer service, communication, and teamwork through a versatile crew-member role at McDonald’s.

Skills: Customer Satisfaction, Customer Service, Communication, Teamwork

AI Engineer for Data Scientists

DataCamp

Feb 2026

AI Engineer for Data Scientists credential from DataCamp.

Generative AI Engineering

IBM

Generative AI Engineering credential from IBM, with skills spanning language models, agents, and applied AI development.

Skills: Prompt Engineering, Data Analysis, SciPy, Project Engineering, Pandas, Agentic AI Development, Project Management, AI Agents, Keras, Tokenization, TensorFlow, Word Embeddings, LangChain, Reinforcement Learning, PPO, Deep Learning, PyTorch, Artificial Intelligence, Jupyter, Fine-tuning, Retrieval-Augmented Generation, Model Pruning, Large Language Models, Flask

ENSA Khouribga

Final year · Morocco

Final-year Data Science & AI Engineering student at ENSA Khouribga, Morocco.

• Seeking an end-of-studies internship (PFE) in AI Engineering, Generative AI or Data Science.