AI Engineer · Research Engineer · PhD, IMT Atlantique

IbuildAIsystemsthatgobeyonddemos.

Agentic AI · LLMs · RAG · Reinforcement Learning

I design, evaluate, and engineer AI systems across multi-agent LLM workflows, retrieval, post-training, reinforcement learning, and applied machine learning: from experiments to APIs, observability, and deployment.

Based in France · Open to AI Engineering, Research Engineering & selected consulting opportunities

Explore my work GitHub LinkedIn

AI Systems Console

live
Data / DocsRetrievalAgent WorkflowEvaluationDeploymentRL / Policy Loop
TRACEretrieval → synth → critique
GROUNDING✓ validated
PhDIMT Atlantique · Lab-STICC
R&DMitsubishi Electric
AI SystemsAgents · LLMs · RAG · RL
Research + Open SourcePeer-reviewed work · reproducible toolkits

Selected Systems

EngineeringAIsystems,notjustprototypes.

A selection of systems where the interesting part is not only the model, but the architecture, evaluation, reliability, and deployment around it.

Flagship System

DocuMind AI

Multi-Agent Document Intelligence Platform

A supervisor-led multi-agent document system coordinating retrieval, summarization, LLM generation, critique, and bounded corrective retrieval over grounded evidence.

DocuMind multi-agent architecture: a Supervisor routes a user query to a Retriever Agent (hybrid dense and BM25 retrieval, RRF fusion, cross-encoder reranking), optionally a Summarizer Agent, and a Synthesizer Agent that performs LLM generation over the retrieved evidence. A Critique Agent checks whether the answer is grounded; if not, a bounded corrective retry reformulates the query and routes back through the Retriever Agent before generation is attempted again.bounded corrective retryUser QuerySupervisor/ RouterRetrieverAgentDense+BM25RRF→RerankSummarizerAgentSynthesizer AgentLLM GenerationCritique Agentgrounded?AnswergroundedReformulatecorrective

137-question evaluation

Result

+6.2

MRR points vs dense-only retrieval

Result

+5.0

nDCG@10 vs dense-only

Eval fact

0.96–0.99

Faithfulness

Eval fact

92.9%

Correct abstention

LangGraphHybrid RAGDense + BM25
Case Study GitHub

LLM Post-Training

MailTune

LLM Post-Training & Evaluation System

A QLoRA fine-tuning project built to test whether a small open LLM can learn concise, casual email style while preserving facts, with controlled baselines and an evaluation suite rather than a single before/after example.

Qwen2.5-1.5B-InstructQLoRAPEFT

Illustrative style example

Before fine-tuning

Dear Sarah, I hope this message finds you well. I wanted to reach out regarding the project timeline...

QLoRA fine-tuned

Hi Sarah, quick update on the timeline: here's where things stand.

Result

73%

win rate vs prompted baseline

Result

44% → 6%

fabrication rate (base → QLoRA)

Config

0.28%

of parameters trained (4.36M)

See Evaluation GitHub

Secondary System

RepeatRadar

Semantic Repetition Detection for Long Documents

A document-analysis system designed during PhD writing to find semantically repeated ideas across chapters and suggest controlled rewrites.

GeminiFAISSFastAPI

Illustrative repetition example

Chapter 2

The evaluation protocol separates retrieval quality from final answer quality so that improvements can be attributed to the correct stage of the pipeline.

Chapter 4

Retrieval and generation are assessed independently, allowing changes in answer quality to be traced back to the stage that produced them.

Suggested fix

RewriteShortenCross-reference

Review decision

AcceptRejectEdit
Case Study GitHub

Independent project · Chrome extension

ApplyPilot

Job Application Automation Chrome Extension

Automates LinkedIn Easy Apply with multi-resume matching, deterministic form handling, recovery controls, and optional Claude fallback for questions the local rules cannot answer.

Automation flow

JOB↓
FILTER↓

RESUME MATCH

weighted local matching

keywords 3×CV text 2×filename 1×
↓FORM↓
KNOWN?
RULE/CLAUDE
↓
SUBMIT ✓

CV parsing + matching stay local

Local PDF parsingWeighted resume matchingDeterministic-firstOptional Claude fallback
Case Study

Reinforcement Learning

Deep RL Agents for Super Mario Bros

DQN · DDQN · PPO

A comparative reinforcement-learning project implementing and evaluating value-based and policy-gradient agents in the same environment.

Reinforcement learning loop between the agent (DQN, Double DQN, or PPO) and the Super Mario Bros environment, exchanging state, action, and reward.actionstate, rewardAgentDQN / DDQN / PPOEnvironmentSuper Mario Bros

Observation preprocessing

240×256×3 RGB→skip + max-pool→84×84 grayscale→stack ×4 → (4,84,84)
Experiment
DQN · DDQN · PPO  ·  1,000 episodes  ·  7 actions
View Project GitHub

Capabilities

WhereIworkacrosstheAIstack.

From model behavior and retrieval to APIs, evaluation, and deployment.

Agentic AI Systems

Stateful workflows, tool calling, routing, memory, human-in-the-loop control, corrective retries.

LangGraph · Tool Calling

Retrieval & RAG

Hybrid dense/BM25 retrieval, RRF fusion, cross-encoder reranking, citation grounding.

Qdrant · RRF Fusion

LLM Post-Training & Evaluation

QLoRA/PEFT fine-tuning, dataset construction, baselines, pairwise LLM judging.

QLoRA · LLM-as-Judge

Reinforcement Learning

Value-based and policy-gradient agents, multi-agent RL, reward design, exploration.

DQN · PPO

ML / Computer Vision

CNNs, LSTMs, tracking, pose estimation, temporal modelling, classical ML.

PyTorch · OpenCV

Production Engineering

Async APIs, data stores, containerization, CI, and observability.

FastAPI · Docker

Research & Reinforcement Learning

Researchthatbecameworkingsystems.

My background is not limited to application-layer AI. My PhD work focused on algorithm design, simulation, benchmarking, and open-source implementations for telecommunications systems.

Mar 2022 – Sep 2022 · Rennes, France

Machine Learning R&D Engineer: Deep Reinforcement Learning

Mitsubishi Electric R&D Centre Europe

Nov 2022 – Dec 2025 · Brest, France

Doctoral Researcher: Algorithm Design & Optimisation

IMT Atlantique · Lab-STICC (CNRS UMR 6285)

Developed a CTDE multi-agent actor-critic system in PyTorch for real-time beamforming optimisation in next-generation MIMO networks.

  • · Centralized training with decentralized execution under partial observability
  • · Reward function design and training pipeline built from scratch
  • · Designed for decentralized real-time inference

Thesis: “Distance-Centric Design Methods for Parallel Concatenated Convolutional Codes”

  • · Novel decoder-free optimisation / minimum-distance estimation method
  • · Full cycle: formulation → implementation → simulation → benchmarking → publication

~10%

Outage-probability reduction vs. brute-force baseline

Centralized training with decentralized execution: a global critic branches into three agents, each consuming a local observation and producing a decentralized policy.Global CriticAgent 1local obsAgent 2local obsAgent Nlocal obsdecentralized policiescentralized trainingdecentralized inference
Interleaved code frame decomposed into a search graph, with a benchmark arrow showing a decoder-free estimation method compared against a slower reference baseline.interleaved frame decompositiondecoder-free path searchbenchmark

~100×

runtime speedup over referenced state-of-the-art baselines

2 peer-reviewed publications · 1 submitted manuscript · 3 open-source toolkits

Publications

  1. 01

    A Novel Parallel Concatenated Convolutional Code Structure Based on Frame Decomposition

    ISTC 2025, Los Angeles

  2. 02

    Efficient Decoder-Free Minimum Distance Estimation for Concatenated Convolutional Codes

    IEEE VTC 2025-Spring, Oslo

  3. 03

    Distance-Centric Joint Interleaver and Structural Code Design for Concatenated Convolutional Codes

    IEEE OJ-COMS · submitted, 2026

Research Toolkits

DPCCC Toolkit

Decomposed Parallel Concatenated Convolutional Codes

Open-source C toolkit for the design and simulation of Decomposed Parallel Concatenated Convolutional Codes, including decoder-free interleaver design and Monte Carlo BER/FER simulation.

CTurbo CodesARP Interleavers
GitHub

mHDEstimator

Decoder-Free Minimum Distance Estimation

Decoder-independent minimum Hamming distance estimator for LTE turbo codes using Return-To-Zero sequence detection and graph-based search.

C99QPP InterleaversRTZ Detection
GitHub

ARP Interleaver Designer

Almost Regular Permutation Interleaver Search

C tool that searches Almost Regular Permutation interleavers using decoder-free minimum-distance estimation.

CRSC EncodersLTE Block Sizes
GitHub

Additional Work

SelectedprojectsacrosstheAIstack.

Video Anomaly & Weapon Detection

Computer-vision pipeline combining object detection, multi-object tracking, pose estimation, and spatio-temporal modelling for behavioural anomaly and weapon detection in video streams, built as freelance client work.

YOLO11ByteTrackPose Estimation
Private / client work, no public repo

Video Violence Detection

ConvLSTM-based video classification system for violence detection, evaluated on ~2,000 real-world videos with a Flask interface.

TensorFlowConvLSTMFlask93% accuracy, ~2,000 videos
GitHub

Research Document Q&A RAG

Lightweight RAG pipeline for querying PDF research papers in natural language with grounded source retrieval.

LangChainFAISSGemini
GitHub

Face Emotion Detection

Real-time facial emotion classification from webcam input using a CNN trained on FER-2013 and OpenCV face detection.

CNNTensorFlow/KerasOpenCV
GitHub

Face Recognition

Real-time face recognition pipeline using OpenCV Haar-cascade detection and LBPH recognition.

OpenCVHaar CascadeLBPH
GitHub

University FAQ Chatbot

Supervised project

Supervised a final-year Computer Science dissertation project building an NLP FAQ assistant using intent classification, T5 paraphrase augmentation, and a curated university-services dataset.

scikit-learnT5NLP
GitHub

Experience

Fromresearchtoproduction.

2023 – Present

Applied AI & Machine Learning Engineer

Freelance / Client Projects · Remote

End-to-end AI system delivery across LLM, ML, automation, and computer vision, from requirements to deployment.

Nov 2022 – Dec 2025

Doctoral Researcher: Algorithm Design & Optimisation

IMT Atlantique, Lab-STICC (CNRS UMR 6285)

Decoder-free optimisation algorithm with a ~100× runtime speedup over the prior approach.

Mar 2022 – Sep 2022

Machine Learning R&D Engineer: Deep Reinforcement Learning

Mitsubishi Electric R&D Centre Europe, Rennes, France

Multi-agent CTDE actor-critic system in PyTorch for MIMO beamforming; ~10% outage-probability reduction.

Education

PhD, Telecommunications Engineering

IMT Atlantique, Brest, France

2022 – 2025

MSc, Machine Vision and Artificial Intelligence

Université Paris-Saclay, France

2021 – 2022

MEng, Telecommunications Engineering

Lebanese University

2017 – 2022

BSc, Computer Science

Lebanese University

2018 – 2021

About

Buildit.Measureit.Improveit.

I like AI work where the difficult part begins after the first demo: deciding what to measure, understanding why a system fails, and putting reliable engineering around probabilistic components. I approach AI engineering the same way I approach research: build, benchmark, inspect failure modes, and iterate.

PhD, IMT Atlantique · Former ML R&D, Mitsubishi Electric

Let'sbuildsomethingthathastowork.

Open to AI engineering, research engineering, agentic AI / LLM systems, reinforcement learning, and selected consulting engagements.

Email me LinkedIn GitHub Download CV

Brest, France