Volunteer Research Assistant at MBZUAI · MSc Thesis at CentraleSupélec

Yacine Baghli

I build generative models and agentic AI systems from scratch.
Currently working on discrete diffusion for structured sequence reconstruction
and single-cell computational biology for autoimmune diseases.

Volunteer Research Assistant · MBZUAI (Zhao Lab) · MSc Thesis at CentraleSupélec / IETR (CNRS)

GitHub LinkedIn Get in Touch

Recent Activity

Jul 2026🏆 1st Place at OpenAI Build Week Hackathon in Paris — Built Mimex, a learn-by-demonstration agent for Codex. Show it once, it learns, it does.
Jul 2026Built ClearBorder at RAISE Summit × Google DeepMind Hackathon — multi-agent customs clearance with live translation, persistent case memory, and computer-use portal navigation
Jun 2026Joined MBZUAI (Zhao Lab) as Research Assistant — building a T-cell single-cell atlas for Multiple Sclerosis
Jun 2026Released Bio-WM — biological world model with DishBrain-inspired spiking neural network
Jun 2026TraceDiff achieves state-of-the-art on the Clustered Nanopore Reads benchmark — manuscript in preparation
Apr 2026Released SynthDNA-Lab — open-source synthetic DNA pipeline
Apr 2026Released MiniLM-Bench — 8 attention mechanisms from scratch
May 2026Built KAIROS at SOPHiA Genetics × ETH Zurich AI for Health Hackathon
Mar 2026Built Embodied Agent at HACK 2026 — ETH Zurich Robotics Club
Mar 2026🥈 2nd Place at the EPFL Quantum Hackathon — ML & Quantum ML for option pricing

Research

Master's Thesis · CentraleSupélec / IETR (UMR CNRS 6164)

TraceDiff

Discrete Diffusion Models for Structured Noise Understanding

A discrete diffusion model that learns the structure of noise — insertions, deletions, and substitutions — rather than treating corruption as generic. Built an alignment-marginalizing encoder with attention pooling over traces and alignment hypotheses, enabling robust fusion of noisy observations. Achieves state-of-the-art on the Clustered Nanopore Reads benchmark, outperforming prior methods (TrellisBMA) on edit distance and failure rate while maintaining constant-time inference. Full training stack built from scratch: mixed-precision (BF16), curriculum noise scheduling, systematic ablation studies across 5 architectural dimensions.

Discrete DiffusionAxial AttentionStructured NoiseSOTAPyTorchBF16

First-author manuscript in preparation — targeting workshop submission (NeurIPS/ICML)

Volunteer Research Assistant · MBZUAI (Zhao Lab)

T-Cell Single-Cell Atlas for Multiple Sclerosis

Single-Cell Computational Biology for Autoimmune Diseases

Building a T-cell atlas for Multiple Sclerosis by integrating longitudinal scRNA-seq mouse datasets (EAE model) with public human MS datasets. Developing computational pipelines for single-cell data analysis: quality control, batch correction, clustering (Leiden), differential expression, and trajectory analysis. Collaborating with immunologists on clinically grounded ML — translating computational findings into experimentally testable hypotheses about T-cell subpopulations and immune checkpoint (VISTA) dynamics.

scRNA-seqScanpyTrajectory AnalysisAutoimmuneClinical ML

Ongoing — Mohamed bin Zayed University of AI, Abu Dhabi

Open-Source Tool

SynthDNA-Lab

Multi-Vendor Synthetic Dataset Generator

Production pipeline generating datasets calibrated to empirical noise profiles from 7 synthesis platforms and 3 sequencing backends (Illumina, Nanopore, PacBio). 3-layer error injection: synthesis → PCR → sequencing.

Noise ModelingIlluminaNanoporePacBioMIT License
View on GitHub →

Selected Projects

🏆 1st Place · OpenAI Build Week Paris

Mimex

A learn-by-demonstration agent for OpenAI Codex. Record yourself completing a task on any website while talking through it — Mimex turns the recording into a skill file, then uses that skill to execute the task autonomously. Voice instructions layer intent on top of action, so the agent understands the why, not just the what. Shipped to Ginse (agent app store).

OpenAI CodexAgentic AILearn-by-DemoVoice-to-Intent1st Place
GitHub →
Bio-Inspired AI

Bio-WM

A research prototype world model bridging digital perception and biological computation. CNN encoder coupled with a DishBrain-inspired spiking neural network (LIF neurons, STDP plasticity) for bio-inspired sequential decision-making.

World ModelsSpiking NetworksSTDPBio-Inspired
GitHub →
Multi-Agent AI · RAISE Summit × Google DeepMind

ClearBorder

A persistent multi-agent customs clearance system using 3 chained Gemini primitives: live two-way supplier calls, persistent case memory across sessions, and computer-use portal navigation — with a human-approval gate before submission.

Multi-AgentGeminiComputer UsePersistent Memory
GitHub →
Architecture Benchmarking

MiniLM-Bench

8 Transformer attention mechanisms (MHA, GQA, MQA, SWA, DiffAttn, MLA, MoH, NSA) implemented from scratch. Controlled pre-training study on FineWeb-Edu with profiling infrastructure and 42 unit tests.

TransformersPyTorchFrom Scratch
GitHub →
World Models · Hierarchical Planning

H-LeWM

Hierarchical Latent World Model with micro/macro temporal scales for long-horizon planning. JEPA-based latent predictor with hierarchical CEM planner reducing autoregressive drift. Tested on PushT and Two Rooms environments.

JEPAWorld ModelsHierarchical PlanningCEM
Multimodal AI · ETH Zurich

Embodied Agent

VLM + SLAM spatial mapping + cross-session knowledge persistence. Spatial-semantic representations stored and retrieved across sessions for transfer learning. HACK 2026 at ETH Zurich Robotics Club. Selected from 500+ applicants.

VLMsSLAMEmbodied AIMemory
GitHub →
Agentic AI · ETH Zurich

KAIROS

LLM-driven multi-agent clinical platform with automated API orchestration across NCBI, PubMed, and ClinVar. Real-time variant annotation, evidence retrieval, and pathogenicity scoring. SOPHiA Genetics × ETH.

LLM OrchestrationMulti-AgentGenomicsFastAPI
GitHub →
LLM Evaluation · Google DeepMind

Reasoning Benchmarks

Synthetic evaluation benchmarks targeting failure modes in LLM generalization and robustness under distribution shift.

LLM EvaluationBenchmarking
🥈 2nd Place · EPFL

Quantum ML for Finance

Built from scratch both classical (LSTM) and Quantum ML models for option pricing — implementing, training, and evaluating each pipeline end-to-end. Also worked on stock prediction using Quantum Reservoir Computing (QRC). EPFL Quantum Hackathon, Mar 2026.

Option PricingQRCLSTMQuantum MLEPFL

Experience

Jun 2026 — Present

Volunteer Research Assistant — MBZUAI (Zhao Lab)

Mohamed bin Zayed University of AI · Abu Dhabi · Volunteer · Remote

Building a T-cell single-cell atlas for Multiple Sclerosis. Integrating longitudinal scRNA-seq mouse datasets with public human MS data. Developing computational pipelines (Scanpy) for clustering, trajectory analysis, and differential expression. Collaborating with immunologists on clinically grounded ML for autoimmune disease biomarker discovery.

scRNA-seqScanpyClinical MLAutoimmuneComputational Biology
Feb 2026 — Present

Master's Thesis — CentraleSupélec / IETR (UMR CNRS 6164)

Rennes · 6 months

Discrete diffusion models for structured noise understanding. Designed TraceDiff — an architecture that learns the structure of noise (insertions, deletions, substitutions) rather than treating corruption as generic. Alignment-marginalizing encoder with attention pooling. Achieved state-of-the-art on the Clustered Nanopore Reads benchmark. Full training stack from scratch: mixed-precision (BF16), curriculum scheduling, ablation studies.

Discrete DiffusionSOTATransformersPyTorchBF16Ablation Studies
2025 — 2026

ML Engineer — Est Ouvrages

Apprenticeship · 12 months

Designed LLM evaluation frameworks for reasoning consistency and robustness. Built end-to-end ML pipelines for structured data extraction.

LLM EvaluationPythonSQLCI/CD
Jun — Jul 2024

Research Intern — GREEN Laboratory (CNRS)

Université de Lorraine · Nancy · 2 months

Implemented a Texas Instruments C2000 F28379D microcontroller board (GPIO, ADC, PWM, ISR). Real-time BLDC motor control programming in C using Code Composer Studio. Communication setup for data transmission to a local web server.

C/C++Embedded SystemsMicrocontrollersReal-Time Control
Sep 2023 — May 2024

Full-Stack Developer — Eco Motion Team

Shell Eco-Marathon · 9 months

Full-stack development of a website for live vehicle tracking. Designed an energy meter and sent consumption data from an ESP32 (C) via fiber optic to a smartphone. Developed an Android HMI to send/receive real-time consumption and GPS data to a MySQL database. Designed the HMI holder in 3D using Fusion 360.

AndroidMySQLESP32CReal-TimeFusion 360
Jan — May 2024

Mathematics Tutor — Polytech Nancy

Teaching · 5 months

Provided math tutoring sessions for students with two years of post-secondary education. Led personalized support sessions and designed exercise sets to help deepen understanding of mathematical concepts.

TeachingMathematicsMentoring

Publications

2026

Discrete Diffusion Models for Structured Noise Understanding in Sequential Data

Y. Baghli, B. Tegin · In preparation — targeting workshop submission (NeurIPS/ICML)

Open Source

SynthDNA-Lab

Multi-vendor synthetic DNA dataset generator

Python · MIT

Bio-WM

Biological world model — DishBrain-inspired SNN

Python

MiniLM-Bench

8 attention mechanisms benchmarked from scratch

PyTorch

GenomicSophia / KAIROS

LLM-driven multi-agent genomic triage platform

Python · FastAPI

HACK 2026 — SlamAss

Multimodal embodied agent with cross-session memory

Python · Team project

Education

2023 — 2026

Polytech Nancy — Université de Lorraine

MSc Engineering — Intelligent & Autonomous Systems

Thesis at CentraleSupélec (IETR, UMR CNRS 6164)

2024

ÉTS Montréal

Exchange — Computer Vision & Signal Processing

2021 — 2023

Université de Lorraine

CPGE — Mathematics & Computer Science

Technical

Research

Diffusion ModelsTransformersWorld ModelsAgentic AISelf-Supervised LearningRepresentation LearningFoundation ModelsMultimodal LearningSpiking Networks

Stack

PythonPyTorchCUDAC/C++BF16/FP16Multi-GPUScanpyGitLinux

Methods

Ablation StudiesControlled ExperimentsReproducible ResearchClinical MLOpen Source