Projects

Selected work

Five projects that show the same idea applied across modalities: learn a generative model of something expensive to simulate, then use it. Where the work is public, the paper, dataset and code are linked.

01Language2025–2026

Kumatigi — French↔Bambara translation corpus

A 67k quality-scored parallel corpus for a language with almost no machine-readable data, released openly and published at Findings of ACL 2026.

Bambara is spoken by millions across Mali and the wider Mande region, yet reliable parallel data barely exists. I built a curation-plus-augmentation pipeline that assigns a quality score to every sentence pair, then used round-trip translation for fluency and back-translation for vocabulary coverage. Handling Bambara’s orthographic variability turned out to matter more than raw corpus size: filtering on quality gave +3–4 BLEU over volume-matched baselines, confirmed by human adequacy and fluency ratings. Sole-authored, and released as versioned datasets on the Hugging Face Hub.

Paper Dataset

02Vision2021–2025

Generative radio maps for indoor Wi-Fi planning

cGAN, diffusion and physics-informed ResUNet models that predict Wi-Fi coverage from a floor plan in place of ray-tracing simulation. Patent filed; Top-5 at the ICASSP 2025 Grand Challenge.

Planning a Wi-Fi deployment means evaluating thousands of candidate access-point layouts, and each evaluation traditionally needs a ray-tracing simulation too slow to sit inside an optimisation loop. I reframed radio-map prediction as image-to-image generation: IRGAN (IFIP Networking 2023) predicts a full coverage map from a single 2D floor plan at 2.5 dB RMSE, fast enough to be called repeatedly. Later work added conditional diffusion and a physics-informed ResUNet that generalises to unseen buildings and frequencies — Top-5 of 16+ teams at ICASSP 2025. A separate line of work closed the sim-to-real gap using crowd-sourced 5 GHz measurements from live home networks (EuCAP 2024). The learned model then drives metaheuristic optimisation for automatic access-point placement, with a patent filed on propagating model uncertainty through that optimisation.

Paper

03Sequences2025–

Synthetic mobility traces with generative transformers

Transformer-based generative models that produce realistic human-mobility trajectories, so operator data can be exploited without exposing real subscribers.

Current work at Orange Research. Mobility data derived from network signalling is enormously valuable for transport planning and urban analytics, but it is also among the most sensitive data an operator holds. The goal is generative models whose samples preserve the statistical structure that makes the data useful — trip distributions, dwell times, spatial correlation — while corresponding to no real individual. Ongoing; details constrained by confidentiality.

04Vision2021–2022

Satellite image time-series completion

A fusion network that uses Sentinel-3 hyperspectral data to reconstruct missing dates in Sentinel-2 multispectral series, for coastal water monitoring.

Optical satellite time series are full of holes — clouds, revisit gaps — exactly when continuous observation matters most. Built at LISIC, this architecture trades on the complementary strengths of two Copernicus missions: Sentinel-2 has the spatial detail, Sentinel-3 the revisit frequency and spectral richness. Fusing them reconstructs plausible imagery for dates where Sentinel-2 has none. Published at ICASSP 2022 and GRETSI 2022.

Paper

05Language2025–2026

Transformers from scratch — NMT, sentiment, RAG

Open-source implementations built from first principles: a full seq2seq Transformer for translation, an encoder for sentiment classification, and a RAG pipeline over a vector database.

Written to keep the fundamentals sharp and to have something readable to point students and interns at — attention, positional encoding, masking and the training loop implemented directly in PyTorch rather than imported. The RAG project covers chunking strategy, hybrid keyword/semantic retrieval, query transformation and evaluation.

Code