About me

I am a Machine Learning Engineer with a Master’s degree (M2) in Mathematics, Vision & Learning (MVA) from École Normale Supérieure Paris-Saclay.
My work spans ML engineering, deep learning, multimodal modeling, reinforcement learning, and LLM/VLM research, with hands-on experience both in academic labs and applied ML roles.

I have previously worked as:

• Research Engineer, L2TI Lab, Université Sorbonne Paris-Nord, wroked on multimodal object detection, few-shot learning, transformer-based models, and dataset creation.
• ML Engineer & Data Analyst, United Nations programs: applied ML, image analysis.
• Research Intern, ISIR Lab (Sorbonne University): supervised and RL fine-tuning for LLMs/VLMs, exploration dynamics in RL for LLMs, and value modeling for planning for robots.

My research includes a peer-reviewed journal publication, an ICIP 2024 oral presentation, and an arXiv preprint.

Open Source

1. Unified sequence parallelism implementation for HF diffusers

PR link

2. Vectorized IoU computation for PerceptionMetrics library

PR link

Selected Projects

1. MVA Research Internship: RL + LLM/VLMs for agentic environments

Explored:

• LLM/VLM bias effects on exploration
• LLM/VLM priors as exploration drivers
• Enhanced value approximation via local utility functions for VLMs Report LLM finetuning VLM finetuning framework

2. Real-time Meeting Copilot

A production-style meeting assistant combining speech-to-text inference, LLM inference, and vector search for live transcription and interactive Q&A.
code

3. Minimalist voice-assistant implementation:

A minimialist implementation for a voice AI chatbot with low latency, and some simple tool calling implementaiton for LLM. code

4. Curiosity-Driven RL for LLMs

Designed a curiosity-based exploration framework using action-level novelty and sequence-level novelty via a T5-based temporal predictor.
code

5. PnP-Flow: Plug-and-Play Image Restoration with Flow Matching

Trained a 2D Flow Matching model and integrated it into a plug-and-play restoration algorithm.
code · report

6. Fine-tuning LLaMA-2 on Personal Chats

End-to-end pipeline for fine-tuning LLaMA-2 on user-specific chat data using free Colab GPU.
code

Publications

1. Indirect Attention: Turning Context Misalignment into a Feature B. Bahaduri, H. Talaoubrid, F. Feng, Z. Ming, A. Mokraoui
   paper

2. A Comparative Attention Framework for Better Few-shot Object Detection on Aerial Images P. Le Jeune, B. Bahaduri, A. Mokraoui
   Pattern Recognition 2024
   paper · code

3. Multimodal Transformer using Cross-Channel Attention for Object Detection in Remote Sensing Images
   B. Bahaduri, Z. Ming, F. Feng, A. Mokraoui
   ICIP 2024 (oral)
   paper · code