Technical writing.
Lessons from architecture, integration or UI pattern decisions. Drawn from professional and personal projects.
Multi-datasource Spring Boot: 3 databases, 1 transaction
On Evelity, back-office editing, mobile API reads and analytics run on three distinct PostgreSQL databases, with atomic transactions that still hold across them.
Visual SLAM ↔ Python: a clean bridge via subprocess
Integrating a C++ Visual SLAM engine into a FastAPI back-end without exotic bindings, then aligning the result on a floor plan with only two GPS constraints.
AI matcher: blending a deterministic score with an LLM verdict
How ApplyDesk scores 100 offers in a fraction of a second in the browser, and keeps the LLM budget reasonable for a reasoned verdict only on those that deserve it.
Computer vision · 2018-2020
Four papers from my research master's at the University of La Rochelle (MIA Laboratory). Detailed summaries here; PDFs are not hosted (IEEE copyright).
Colorizing genuine archival photos: palette + scribbles
CNN colorization networks are trained on colour photos converted to grayscale, which makes them ill-suited to genuine archival B&W (silver halide). I improve the output with two hints: an automated global palette + manual scribbles where it matters (flags, monuments).
Colorizing a B&W movie with a salient-colour palette
A palette of 'salient' scene colours (distinct from the 'memorable' ones the network already infers well) is injected into the CNN as an additional input. One palette covers hundreds of frames, which suits movie colorization without per-frame scribbling.
Detecting handwriting anomalies via a convolutional autoencoder
A convolutional autoencoder that detects, without annotation, that a portion of text was written by another hand. Tile splitting + partial Radon transform. Discrimination emerges during training via batch-to-batch shuffling.
Visible + NIR information for CNN face recognition
An original VNIR dataset (52 identities, 3 poses) captured by removing the ICF of a consumer camera. Two CNNs compared. Finding: 3-channel full-spectrum beats 4-channel RGB+I, and the blue channel benefits the most from NIR.