Pierre – Python, AI agent development, LLM

An open-source Python framework for real-time acquisition and processing of biosignals. It supports researchers and engineers building brain-computer interfaces and physiological monitoring applications with a modular pipeline for EEG, ECG, EDA, eye tracking, and other biosignals. The framework has been used in production by neurotechnology companies and research institutions worldwide.

• Authored and maintained the framework from its inception, leading all architectural decisions;
• Designed a modular software architecture based on execution graphs and a plugin system;
• Implemented low-latency asynchronous processing for real-time biosignal handling;
• Integrated hardware drivers for a wide range of devices covering EEG, ECG, EDA, and eye tracking;
• Implemented real-time communication protocols including LSL, ZeroMQ, and OSC;
• Built signal processing and machine learning components based on applied neuroscience, including BCI paradigms and EEG neurophysiology;
• Designed user interfaces for both web-based control and stimulus presentation;
• Developed the ASAP pipeline (2022), based on Bayesian accumulation of Riemannian probabilities, which significantly outperformed standard P300 classification methods on public datasets;
• Contributed to the StAR paradigm (2024), which raised cVEP accuracy from 65.6% to 96.3% with 88 seconds of calibration and reached 97.5% online on a dry 8-electrode EEG, paving the way for BCIs usable outside the laboratory;
• Grew and supported the open-source community around the project;
• Delivered more than 50 talks and workshops on these topics in both academic and industry settings.

An AI-powered due diligence platform for M&A teams working with large document collections such as data rooms, financial reports, and legal contracts. It lets analysts query thousands of heterogeneous documents in natural language and receive answers with full source traceability, reducing the manual review effort typically required during transaction analysis.

• Designed and developed the end-to-end system architecture;
• Built a multi-format ingestion pipeline (PDF, DOCX, XLSX, images) with OCR, semantic chunking, and structured metadata extraction;
• Implemented hybrid indexing combining dense vector search with a knowledge graph layer for entity resolution and multi-hop reasoning;
• Enforced citation traceability linking every generated answer back to its exact source (page, paragraph, document);
• Implemented hallucination detection through post-generation verification using NLI and RAGAS metrics, along with a cite-or-abstain policy;
• Built agentic orchestration for complex queries that required decomposition and cross-document synthesis;
• Packaged and delivered the product as a deployment-ready Docker Compose stack;
• Produced detailed technical architecture documentation, integration guides, and volume-based cost scenarios covering indexing, inference, and storage.

A demonstration project showing how open-weight large language models can be fine-tuned to reduce excessive censorship and bias while preserving general-purpose assistant capabilities. The work focused on Chinese LLMs and covered a full workflow: parameter-efficient fine-tuning, preference-based training data, and automated evaluation on modest hardware. It is aimed at researchers and engineers interested in alignment, model behavior modification, and lightweight fine-tuning techniques.

• Designed and implemented an end-to-end fine-tuning pipeline in a reproducible Jupyter notebook;
• Applied Low-Rank Adaptation (LoRA) as a parameter-efficient fine-tuning method, avoiding the cost of full model retraining;
• Fine-tuned the Qwen2.5-0.5B-Instruct model on the nbeerbower/GreatFirewall-DPO preference dataset, covering both English and Chinese queries;
• Built a custom evaluation harness using GPT-4.1-mini as a less biased judge to score responses on a curated set of sensitive prompts;
• Compared base-model and fine-tuned outputs to quantify the reduction in self-censorship while monitoring for capability regression;
• Documented the methodology, its limitations, and the rationale for choosing a small model for the demonstration, with the approach designed to generalize to larger models in the Qwen family and beyond;
• Published the adapter weights, the merged fine-tuned model, and the full notebook as an open-source reference.

An interactive web platform for exploring Intangible Cultural Heritage elements through semantic relationships and visual navigation. It enables users to browse several hundred heritage elements from more than 100 countries and discover connections across domains, themes, geography, biomes, threats, and Sustainable Development Goals. The platform serves the general public, researchers, States Parties, and the broader heritage community as an open-access knowledge tool.

• Unified data aggregation across multiple heterogeneous sources into a single coherent dataset for the platform;
• Designed the project’s ontologies and integrated them with external ontologies to enable semantic navigation and interconnection of heritage elements;
• Implemented a full RDF triple store layered over MySQL, providing the semantic data backbone of the platform;
• Built indexing interfaces used to curate, structure, and refine the indexing of inscribed elements;
• Defined the conceptual design of the data visualizations, concept only, not the visual implementation;
• Developed the functional frontend of the platform’s interfaces;
• Implemented the caching policy and the generation of JSON graph data consumed by the visualization layer.

A research project exploring transfer learning for ERP-based brain-computer interfaces, with the goal of reducing or eliminating the per-subject calibration phase. The work focused on the P300 evoked potential and on building a generic single-trial classifier that generalizes across subjects and attention tasks. Removing the calibration phase would make ERP-based BCIs more accessible, especially for users with limited attention span.

• Framed the problem as a sequence classification task and chose an LSTM-based deep learning approach over the linear baselines that dominate the field;
• Reproduced the original study’s baseline (downsampling plus regularized LDA, following Blankertz et al. 2011), reaching a grand mean AUROC of 0.61 in line with the published results;
• Built the full data preprocessing pipeline using MNE: bandpass filtering, downsampling, baseline correction, cropping, standardization, and conversion to MNE Raw, Epochs, and Evoked objects;
• Designed and implemented a stacked LSTM architecture (three LSTM layers of 128, 64, and 32 units, followed by a dense layer and sigmoid activation) trained with the Adam optimizer and binary cross-entropy;
• Engineered Keras callbacks for stateful training across full epochs, AUROC tracking on the validation set, model checkpointing, and a custom early stopping policy based on validation AUC;
• Addressed overfitting through 40% input and recurrent dropout, small batch sizes for frequent weight updates, and careful management of statefulness;
• Implemented a leave-one-subject-out cross-validation scheme covering 3 conditions and 13 subjects, training and evaluating 117 models in roughly 53 hours on a Tesla K80 GPU;
• Improved the grand mean AUROC from 0.61 to 0.70, a 0.09 absolute gain over the published baseline, demonstrating that LSTM networks can be successfully applied to single-trial ERP classification across subjects and tasks;
• Produced spatio-temporal scalp topography visualizations and a written report covering methodology, results, limitations, and recommended next steps (full hyperparameter sweep, GRU comparison, larger and standardized ERP datasets, montage-independent estimators).