Christopher Bourez

Development of Scenario's GPU inferences: serverless ML endpoints for image, video, and 3D generation, deployed on Replicate and Modal, from the first commit through 200+ production PRs.

- Diffusion inference at scale (FLUX.1 dev/Kontext/Krea, FLUX.2, Qwen Image, Z-Image), including ControlNet/depth/canny, Redux, inpaint/fill, img2img, negative prompting, and torch compilation for latency reduction.

- Platform migration of inference and training workloads from Replicate (cog) to Modal serverless GPU infrastructure (inference, upscaling, fast-flux, training), cutting cold-start and operational overhead.

- Dynamic multi-model multi-pipeline serving system with LoRA hot-loading/routing (LorasManager over Modal Dict/Volume), async execution, and GPU right-sizing per (resolution, batch, rank); diagnosed and fixed deep VRAM leaks (diffusers hook cycles, torch.compile reference cycles, retry-path pipeline leaks).

- LoRA ecosystem: Implemented Flux.2 Lora Training in Kohya/Musubi-Tuner framework, multi-format LoRA compatibility (Kohya, XLabs, ComfyUI, diffusers), text-encoder LoRA loading/unloading, and conversion utilities (regression fixes across transformers/diffusers upgrades).

- Training pipelines: Shipped LoRA fine-tuning services (Kohya, musubi-tuner) for FLUX, Kontext, Krea, Qwen Image, Z-Image, and FLUX.2, with S3 dataset ingestion, preemption/retry resilience, and idempotent resume on spot interruption.

- Panorama/360° generation: Developed seamless skybox generation and upscaling (circular padding, seamless attention, nadir correction, color/variance-matched tiled refinement with IP-Adapter and ControlNet).

- Creative-tooling products: video subtitles (transcription, translation, styling, Remotion-rendered animated themes) and image/video layer extractor

- Security & reliability: Sandboxed untrusted AI-generated code rendering, hardened webhook delivery with differentiated retry policies, and eliminated recurring OOM/stability failures across GPU services.

Automatic extraction of values from scanned invoices

Automatic analysis of contracts (segmentation into clauses, keyword extraction).

Classification of expense lines

Deduplication of supplier and product data

Chatbot

Particularly with Microsoft technologies (CNTK, ML.Net, ONNX), and ABBYY

Several jobs as freelancer:

FITLY July 2016 – Sept 2016 https://getsmartplate.com

Deep neural network trainings and optimisations. Work on requests' latencies.

Gcloud AppEngine / Storage buckets / Datastore / Python Flask / Stackdriver monitoring. Scalability of GPU with AWS EC2 G2 and nvidia-docker

SELECTIONNIST March 2014 - July 2015

Mobile website and mobile application for Android and IOS with image recognition. Image indexation and recognition technology. Transfer agreement with INRIA. Fullstack JS development. Mobile app development in AngularJS + Cordova. Webservices development in NodeJS. Test strategy and development (Jasmine+Protactor). Modernization + scalability : SI migration to Amazon Webservices. Mount of former website with FUSE/S3FS (step 1). Worldwide acceleration and cache with content delivery network technology. Full migration of the PHP site (step 2) with S3 stream wrapper : no more data on instances' disks. Tablet version of the mobile app thanks to responsive design. Push notifications, Facebook Integration, and Deep Linking / Custom URL Schemes Virtual private cloud instantiation (route table, gateway, security profiles for instances). Continuous development (DevOps) with Chef, Docker, Vagrant and Opsworks. Technical development and management. Release V1 : June 2014 Release V2 : October 22, 2014 Note : Selectionnist is the best start-up in France in 2015 at the IMC Awards http://lentreprise.lexpress.fr/marketing-vente/ebusiness/les-champions-de-l internet-francais-en-2015-sont_1645107.html

TESSI Nov 2013 - March 2014

Study about the digital opportunities and the technologies for corporate companies Technology watch and preparation of business meetings