Industry research
Freeda — AI compliance review for the construction industry
Co-Founder & Chief Scientific Officer — Nov 2024 – present, Freeda, Paris
I own the scientific strategy and the AI roadmap, and lead a team of 3 ML engineers. Every layer of Freeda’s AI stack was shipped by that team under my direction:
- Perception. Up to three generations of plan-understanding models (RF-DETR, YOLO-OBB, SegFormer), all improving on the Roboflow baseline, alongside template matching and SAM 2 / SAM 3 one-click area selection.
- Plan description layer. A structured record of measurements derived from detected shapes, readable by agents and humans alike — one representation that stops every downstream step from re-deriving geometry.
- Annotation tool. In-house app for labelling text and vision data across document folders (PDFs, images), built on CVAT’s vision component.
- Document intelligence. Freeda Assistant — MinerU OCR, contextual retrieval, hybrid search and reranking — lets domain experts search thousands of pages instead of reading them. Smart Selection uses LLM contextualisation of retrieved patterns to capture every text span tied to a concept by style and context.
- Agentic verification. Recipe-based: an agent follows a human-authored recipe as its roadmap and a human is asked only at final sign-off, with skills that write comments and annotations directly into the delivered PDFs.
- Codegen. Autonomous coding agents used for dynamic custom workflows — e.g. customer-facing interactive reports built as mini-apps: single-page HTML emitted from a real React + Tailwind app.
- Prompt optimisation. Automated prompt improvement adapted from a Microsoft method, removing manual prompt iteration as the bottleneck on quality.
- Infra / DevOps. I own the runtime these features live on: frontend, backend, CI/CD, NVIDIA Triton Inference Server and Celery workers — the first working prototype ran at less than EUR 300/month on a single NVIDIA L4 GPU.
Product impact: took the compliance workflows, engine and infrastructure from blank page to production in a few months, cutting manual plan-review time 3× versus market standards; live with 10+ paying customers within 10 months.
Gladia — speech infrastructure at commercial scale
Founding Machine Learning Engineer, Speech — Jan 2023 – May 2024, Paris
Speech-to-text infrastructure sold as a public API to 100+ customers (PAYG and enterprise contracts); Gladia was later acquired by OVH.
- Inference research with a direct margin effect. Rebuilt Whisper for production serving (ONNX, NVIDIA Triton, flame-graph profiling): lower cost per hour of audio at better WER than the market at the time.
- Owned a research area end to end. Speaker diarization and voice-activity detection (pyannote, silero), which unlocked the meeting-transcription use case.
Powder — video & multimodal in a consumer gaming product
Machine Learning Engineer, Video & Multimodal — Jul 2021 – Jan 2023, Paris
- Research shipped to consumers. Event and highlight detection in gameplay streamer video.
- On-device deployment. Game identification in video and spoken-language identification exported to ONNX/CoreML.
- Published what shipped. Multi-annotation attention model for video summarization, CVPR Workshop 2023.
- Research infrastructure. Set up the team’s training and evaluation stack (PyTorch Lightning, Comet/WandB, Gin, CI with reproducible experiments).
Academic research
Doctoral work on affine-invariant image comparison at ENS Paris-Saclay under the supervision of Jean-Michel Morel and Julie Delon:
- Covering the Space of Tilts — SIAM IS 2018 / SIIMS 2018.
- Fast Affine Invariant Image Matching (IMAS) — IPOL 2018.
- Affine invariant image comparison under repetitive structures — ICIP 2018.
- SIFT-AID: boosting SIFT with an affine invariant descriptor based on CNNs — ICIP 2019.
- CNN-assisted coverings in the Space of Tilts (ADI-MAS) — ICIP 2020.
- Robust estimation of local affine maps (LOCATE) — WACV 2020.
- Automatic detection of repeated objects in images — ICIP 2021.
- A CNN cloud detector for panchromatic satellite images — IGARSS 2021.
- Robust validation steps for clip classification — preprint 2022 (Powder).
PhD thesis
I was a PhD Student at the ENS Paris-Saclay under the supervision of Jean-Michel Morel and Julie Delon. The manuscript is available at:
Awards & fellowships
- Laureate, FMJH Doctoral Program — Fondation Mathématique Jacques Hadamard. Three-year competitive fellowship. (2016)
- Laureate, PGSM Masters Scholarship — Paris Graduate School of Mathematics. Two-year international scholarship. (2014)
- Top student, Faculty of Mathematics & Computer Science — Awarded by the vice-chancellor, University of Havana. (2011)
Teaching
- Analyse de Fourier et Hilbertienne — practical lessons, Department of Mathematics, ENS Paris-Saclay. (Sep 2017 – Jan 2019)
- Calcul Différentiel — practical lessons, Department of Mathematics, ENS Paris-Saclay. (Sep 2016 – Jan 2017)
- PDE, Complex Analysis, Mathematics I–III — lectures & practical lessons, Mathematics Department, University of Havana. (2011 – 2014)