Guillaume Jeffroy

Spin-off from LAAS-CNRS, Senscell Biosystems leverages a nanoelectrode biochip technology capable of recording the electrical activity of living neurons in parallel across dozens of channels, akin to a camera sensor. I have been working with the startup since its inception to design and develop the software for exploiting this data.

The project began with a series of mini-POCs to identify the appropriate stack. Each feature is first prototyped, discussed, and validated in isolation before being integrated into the production version. This approach allows for usable software at each stage, reduces integration risks, and enables adjustments as scientific needs become clearer.

The software now covers the entire exploitation workflow and is enhanced with ongoing experiments:

- **Real-time visualization**: Freely navigate the recording like a video player, automatically identifying events.

- **Signal cleaning**: Automatic pre-processing adapted to the specific noise profile of each experiment, with manual adjustment options.

- **Analysis and export**: Summary statistics, figures of merit, graphs, and data exportable to Excel or structured formats for result communication.

- **Traceability**: Each analysis is recorded with its parameters and results, allowing it to be found, replayed on new data, or compared across experiments.

Heavy computations are efficiently handled thanks to optimization efforts and the performance of the chosen language (Rust). Visualization rendering runs in the background to ensure a fluid and responsive interface — a requirement for the tool to be truly usable daily and to help biologists and neuroscientists exploit the data.

The goal is to make signal exploration and processing accessible by using AI agents to translate business requests into complex analyses via dedicated tools; one can describe what they want to do in natural language, and the agent guides the user then launches the right tools (simulation, filtering, event detection, time-frequency analysis).

The project strongly emphasizes performance and interactivity: ClickHouse is used to efficiently store and query large time series, Julia executes heavy computations (signal simulation, multi-band DWT, filtering pipelines, spike detection), and uPlot ensures fast and stable web visualization even with large volumes. The whole is orchestrated via a FastAPI backend with real-time streaming (SSE) and a chat/action interface that allows chaining analyses while maintaining a clear user experience.

Design and development of a platform for mass processing of HR documents (CVs, job descriptions, training sheets, job offers, etc.) in heterogeneous formats (PDF, DOCX, Excel). The system automatically extracts, structures, and harmonizes key information using AI pipelines and parallelized workflows.

The architecture relies on decoupled microservices (FastAPI / Prefect), deployed in containers on French cloud infrastructure (Scaleway), with S3 object storage and a real-time monitoring dashboard. A checkpoint mechanism ensures automatic error recovery, enabling a success rate close to 100% on over 10,000 processed documents.

GDPR compliance is natively ensured by a PII anonymization and detection pipeline integrated into the processing flow.

Python · FastAPI · Prefect · LLM · Sentence Transformers · Docker · Scaleway · React