Selected Work

Research, data engineering, and machine learning systems development.

Deep Learning for 3D Sonar Signal Denoising & Detection

PhD Research • Deep Learning • Applied Mathematics

PhD research work focused on detecting weak underwater acoustic sources at extremely low Signal-to-Noise Ratios (SNR) in passive broadband sonar systems, going from 2D arrays to volumetric 3D (bearing-elevation-time) representations.

1. Simulation & Data Engineering

Designed a fast, and physically accurate sonar simulator from scratch to overcome the lack of real labeled data. The engine models propagation, attenuation, and multipath reflections, generating large-scale synthetic datasets for training.

2. AntoNet Architecture (2D & 3D)

Developed AntoNet, a custom U-Net-inspired CNN tailored for spatial-temporal denoising.

  • Performance: Reached detection gains up to 11 dB, performing extremely close to the theoretical limit of incoherent processing (12 dB).
  • Efficiency: The linear version requires only 1.2k parameters (vs 20.6M for the non-linear), ideal for edge deployments.
  • SOTA Comparison: Reduced Mean Squared Error (MSE) by a factor of >10x compared to existing architectures.
3. Mathematical Analysis & Interpretability

Conducted an analytical study using symbolic regression to characterize the global minima manifold of the network, proving that even a linear AntoNet exhibits non-linear learning dynamics, thus ensuring robustness against Gaussian noise.

Outputs & Publications
  • IPTA 2024: Top-5 paper on AntoNet's linear dynamics. Extended version under review for SPIC Journal.
  • EUSIPCO 2025: Published research on multi-path scenarios and feature-map visualization.
  • Patent Filed: SYSTEME ET PROCEDE DE DEBRUITAGE D'IMAGES 3D, N° 2514119 Covering the full pipeline from signal generation to 3D source tracking.
Tech Stack:
Python PyTorch CUDA Signal Processing Math Modeling
Read more ↓

Self-Hosted LLM Infrastructure & Private AI Server

MLOps • Systems Engineering • GenAI • Homelab

Designed and deployed a private, Linux-based AI server to run state-of-the-art Large Language Models (LLMs) locally. This infrastructure acts as a privacy-first, zero-cost alternative to cloud APIs, securely accessible from anywhere in the world.

1. Hardware Setup & Multi-GPU Inference

The system is built on a headless Linux home server equipped with a dual-GPU setup (RTX 4070 + RTX 3060 12GB). Overcame strict VRAM constraints by distributing model layers across both GPUs via CUDA, maximizing inference throughput for large models.

2. Quantization & Model Optimization

Deployed complex architectures using Ollama and llama.cpp, benchmarking dense models (Llama, Gemma) as well as Mixture of Experts (MoE) architectures. Applied aggressive GGUF quantization techniques (4-bit and 2-bit) to fit high-parameter models into consumer-grade VRAM while minimizing degradation in output quality.

3. Private Web Search & Zero-API Integration

To empower the LLM with real-time internet access without compromising privacy, I deployed a self-hosted web search engine via Docker. This enables Web-Augmented Generation capabilities entirely on-premise, completely bypassing the need for commercial external APIs (e.g., Google/Bing search APIs).

4. User Interface & Zero-Trust Networking

Configured Open-WebUI to provide a ChatGPT-like frontend, enabling fine-grained control over generation parameters. Secured remote access by implementing Tailscale to establish an encrypted Zero-Trust VPN, allowing secure queries from mobile devices globally.

Key Achievements
  • Efficient handling of Multi-GPU memory allocation.
  • Deployed a 100% private, local web-search integration (Zero-API).
  • Built an always-on, fully private global AI assistant architecture.
Tech Stack:
Linux Server Ollama / llama.cpp CUDA / Multi-GPU Self-Hosted Web Search Open-WebUI Tailscale (VPN) Docker
Read more ↓

Speech Impairment Detection

Deep Learning • Audio Processing

Built a classification pipeline using spectrogram preprocessing and CNN/RNN models. Implemented a reproducible pipeline prepared for on-device edge inference.

Full-Stack Technical Translation Platform (DeepL API)

Web Development • Flask • REST API • Server Administration

Developed a self-hosted web application for translating technical documents while preserving their original layout. The app integrates the DeepL API with an advanced custom glossary manager, solving the critical issue of generic translations in highly specialized domains.

Backend Architecture & Database

Built a Python/Flask backend featuring session-based user authentication. Engineered a dynamic routing system for multiple SQLite databases (automatically organized by language pairs, e.g., EN-FR) to store and enforce strict domain terminology during the translation process.

REST API & Server Deployment

Designed RESTful endpoints (CRUD) allowing asynchronous glossary management from the frontend UI. The entire infrastructure is hosted on a personal Linux home server, administered remotely via SSH, ensuring full data privacy and zero cloud hosting costs.

Tech Stack:
Python / Flask SQLite DeepL API RESTful API HTML / JS Linux Server (SSH)