[nom]

[e-mail] | [téléphone] | Fès, Morocco | [lien] | [lien]

Education

Faculty of Sciences and Techniques · Septembre 2024 – June 2026

Master’s in Computer Science | Specialization: Artificial Intelligence and Data Science · Fès, Morocco

Relevant Coursework: Artificial Intelligence, Machine Learning, Deep Learning, Big Data, Cloud Computing, Statistics, Text and Web Mining, Applied Mathematics, Data Analysis, Computer Vision, Business Intelligence.

Professional Experience

SmartlyAI · July 2026 – Present

Voice AI Engineer · Casablanca, Morocco

• Building and shipping production-grade Voice AI Agents using LiveKit, LangGraph, and LangChain, powering real-time conversational voice pipelines deployed to production.
• Designing low-latency, agentic voice workflows integrating speech-to-text, LLM reasoning, and text-to-speech into a unified orchestration layer.
• Developing the supporting backend and agent infrastructure using Node.js, TypeScript, and JavaScript, integrating APIs and real-time services within the voice AI stack.
• Collaborating within a Docker-based deployment workflow to develop, test, and ship voice agent features directly to production environments.

SOFTDIGIS · February 2026 – June 2026

AI/ML Engineer · Fès, Morocco

• Fine-tuned OpenAI’s Whisper Medium model for Moroccan Darija (Arabic dialect) automatic speech recognition using LoRA, achieving 6.2% Word Error Rate on inference — ranking among the best published results on public Darija benchmarks (DODA, Speech to Text by Adiren7).
• Conducted a structured three-stage experimentation pipeline (QLoRA on Kaggle P100 → LoRA on NVIDIA A100), covering dataset engineering, text normalization, and hyperparameter tuning across 138M trainable parameters.
• Deployed the fine-tuned model as a production-ready containerized microservice (FastAPI, Docker) integrated into an end-to-end voice-based address collection system via WhatsApp (Twilio), orchestrated with LangGraph and powered by a Groq-hosted LLM for address validation.

Injaz-Almaghrib · February 2025 – June 2025

Artificial Intelligence Engineer (Generative & Agentic AI) · Fès, Morocco

• Developed an intelligent conversational system powered by a Large Language Model (LLM) using the Gemini API, generating responses exclusively in Moroccan Darija.
• Designed advanced prompting strategies and integrated real-time user interactions via Streamlit, enabling natural com- munication for customer support, product inquiries, and advisory services.
• Implemented dynamic API-based pipelines for content generation, Q&A, and order tracking, significantly improving user engagement and response speed.
• Achievement: Ranked 2 nd place with my team in the regional Injaz-Almaghrib competition.

CGI (Technology & Business Consulting) · April 2024 – August 2024

Software Engineering Intern · Fès, Morocco

• Developed a full-stack web application using Spring Boot, React, JavaScript, HTML, and CSS to measure team creativity and productivity.
• Created an intuitive dashboard for managers to track consultant performance, ticket resolution rates, and break times.

Personal Projects

Flash Attention From First Principles (In Development) · Python, PyTorch, Triton, CUDA

• Currently implementing Flash Attention from first principles, building block-wise matrix multiplication with an online softmax that avoids ever materializing the full attention matrix.
• Explicitly reasoning about the GPU memory hierarchy — fast on-chip SRAM versus slower HBM/DRAM — to minimize memory reads and writes between them.
• Implementing custom fused kernels using Triton and CUDA, with a focus on deeply understanding IO-aware attention and how modern LLM inference engines achieve their speedups.

LLaMA-Style Transformer From Scratch — GQA, KV Cache, RoPE & RMSNorm · Python, PyTorch

• Built a complete LLaMA-style transformer architecture from scratch in PyTorch, implementing all modern components without relying on any high-level libraries or pretrained weights.
• Implemented Grouped Query Attention (GQA) with configurable query and key-value heads, reducing memory bandwidth during inference while preserving model quality compared to standard multi-head attention.
• Designed a KV Cache mechanism for efficient autoregressive generation, enabling [information personnelle]-token decode steps by storing and reusing past key-value projections — eliminating redundant recomputation at each generation step.
• Implemented Rotary Positional Encoding (RoPE) using complex number rotations for relative position awareness, and RMSNorm for stable, lightweight layer normalization as used in modern LLMs.
• Trained on the TinyStories dataset and built a full inference pipeline with prefill and decode phases, demonstrating end-to-end understanding of production LLM inference internals.

RLHF From Scratch — Reward Modeling & PPO Transformer Python, PyTorch, Transformers, RLHF, PPO

• Built transformer-based RLHF architectures from scratch using PyTorch to understand modern LLM alignment pipelines.
• Implemented a Reward Model and a PPO-style Transformer with custom attention mechanisms, embeddings, policy/value heads, and transformer blocks.
• Explored reinforcement learning, reward modeling, and low-level transformer internals with focus on educational and research-oriented implementation.

AI-Powered Social Media Marketing Agent (In Development) Python, LangGraph, LangChain, DALL-E, Meta

Graph API, TikTok API

• Developing a comprehensive multi-agent system to automate creation, publishing, and analysis of social media adver- tisements (Instagram, Facebook, TikTok).
• Implementing a content generation pipeline integrating LLM-based information extraction, advertising copy generation, image creation via DALL-E/Stable Diffusion, and automatic visual composition with Pillow.
• Integrating social media APIs (Meta Graph API, TikTok API) for automated multi-platform publishing with adaptive formatting (1:1, 16:9, 9:16) and time-based scheduling.
• Creating a performance analytics agent collecting metrics (reach, impressions, engagement) and generating insight reports with recommendations based on past performance learning.

Moroccan Darija AI — LLM & ASR Fine-Tuning Python, LoRA, QLoRA, Unsloth, Qwen, Whisper, A100 SXM

GPU

• Fine-tuned large language models and Automatic Speech Recognition (ASR) systems for Moroccan Darija, a low-resource Arabic dialect with limited AI support.
• Applied memory-efficient parameter-efficient fine-tuning techniques (QLoRA with Unsloth) to optimize Qwen for natural language understanding and Whisper for speech recognition.
• Managed the full pipeline: custom dataset engineering, instruction data curation, model optimization, and performance evaluation.

Multi-Agent AI Assistant · Python, LangGraph, LangChain, RAG, SQLite, OpenAI Embeddings

• Designed and built an agentic AI system with three specialized agents orchestrated by LangGraph: a RAG agent for university policy documents, a SQL agent for timetable and faculty data, and an Email agent for professional email drafting and sending via SMTP.
• Implemented persistent conversation memory using LangGraph’s MemorySaver, tool calling with LLM function binding, and a dynamic planner that automatically routes complex multi-step queries.

GPT-2 Small — From Scratch · Python, PyTorch, Kaggle GPU

• Fully reimplemented the GPT-2 Small architecture from scratch in PyTorch, including causal multi-head attention, positional embeddings, and transformer blocks (no pretrained weights used).
• Trained on the TinyStories dataset for approximately 12 hours on a Kaggle P100 GPU over two rounds, achieving a final cross-entropy loss of 1.57 (compared to a random baseline of 10.8).
• Implemented a custom autoregressive decoding loop with temperature sampling to generate coherent story continuations from arbitrary prompts.

AskUniAssistant · Python, RAG, FAISS, OpenAI Embeddings, GPT-4o-mini, Flask

• Developed an intelligent chatbot that allows students to query university regulations in natural language.
• Implemented Retrieval-Augmented Generation (RAG) using FAISS and OpenAI embeddings to index and retrieve information from PDFs, with GPT-4o-mini generating accurate, context-aware responses.
• Built a user-friendly Flask web interface in French for a smooth interactive experience.

LaptopAssistant · Python, Flask, OpenRouter API, LLMs, Prompt Engineering

• Built an intelligent chatbot capable of querying databases in Moroccan Darija, Arabic, and English using natural language.
• Implemented three pipelines (translation, SQL generation, and response formatting) using Flask and the OpenRouter API.
• Applied advanced prompt engineering techniques to ensure reliable, hallucination-free SQL query generation.

English–Arabic Medical Translator · Python, TensorFlow/Keras, Tokenizers, GloVe, Flask

• Developed a web application for English–Arabic medical translation based on an LSTM Seq2Seq (Encoder–Decoder) architecture.
• Implemented embedding layers, LSTM encoder-decoder, and greedy decoding, achieving accurate translations for short to medium-length sentences.
• Integrated a Flask interface for real-time translation.

Image Caption Generator · Python, TensorFlow/Keras, OpenCV, Tokenizers, GloVe, Pandas

• Developed an image captioning system combining Computer Vision and Natural Language Processing to automatically generate descriptive captions.
• Used pretrained CNNs (InceptionV3/VGG16) for feature extraction and LSTM networks for sequence modeling, trained on image-caption paired datasets.
• Implemented image preprocessing, tokenization, and greedy decoding for fluent, contextually accurate captions.

Real-Time AI-Powered Vehicle Analytics System · Python, YOLOv10, OpenCV, NumPy

• Developed a real-time AI-based vehicle monitoring system using YOLOv10 for multi-zone detection, persistent tracking, and live analytics visualization.
• Delivered a professional dashboard with live updates and achieved 99% accuracy in vehicle classification for smart city traffic management and logistics optimization.

Technical Skills

AI & Machine Learning: Generative AI, Agentic AI, RLHF, PPO, Reward Modeling, RAG, LangChain, Transform-ers, LLMs (BERT, GPT, LLaMA), Deep Learning (MLP, CNN, RNN, LSTM), NLP, Computer Vision (YOLO, Vision

Transformers, OpenCV), KV Cache, vLLM, Model Quantization (LoRA, QLoRA)

Programming & Tools: Python (NumPy, Pandas, Matplotlib), SQL, TensorFlow, PyTorch, Scikit-learn, Spark, Hadoop, Git, Docker, Jupyter, Kaggle, Google Colab.

Languages

English (Fluent) | French (Fluent) | Arabic (Native)