Gabriel Kasser

Independent engineering: designing, building and shipping AI-powered products end-to-end (ML, backend, frontend) with full ownership. MTG Trade (mtg-trade.fr) — solo: Computer vision and pricing optimization platform for Magic: The Gathering Stack: 98% card identification accuracy (card + language); 73% with exact edition; ~5s batch inference per 9-card page. Beta launched in 3 months. Max App (max-app.eu): WhatsApp business assistant for craftsmen, LLM-powered quote generation and workflow management. Stack: to balance Datatoy.fr - AI Security Analysis: Agentic tool MCP agents. Stack: Physia (two POCs): Physiological rehabilitation assistant. POC 1: keypoint-driven RNNs for movement analysis, optimized for sub-second classification latency between user action and corrective feedback.

Joined following Datakalab acquisition. Embedded in a research-to-production team focused on on-device ML optimization for Apple hardware.

• • Designed and implemented a production-grade post-training quantization framework targeting ONNX, PyTorch and CoreML — delivering model compression with controlled accuracy degradation for Apple-specific hardware.

• • Built robust packaging and deployment pipelines, significantly reducing friction in the research-to-production handoff.

• • Research on Quantization-Aware Training (QAT) and mixed-precision post-training quantization.

AI startup specialized in neural network compression and on-device optimization. Computer vision systems deployed at Paris CDG and Orly airports, serving millions of passengers annually.

• • Led technical direction across a team of 6 engineers and researchers: clear vision, execution and prioritization under strict regulatory deadlines. Structured the company engineering culture from scratch: clean code practices, CI/CD, agile methodology, infrastructure-as-code, and collaborative Git workflows.

• • Biometric gate - unicity detection: Shipped a person re-identification system for biometric gate control at CDG and Orly airports. Achieved 97% unicity detection accuracy trained on a privacy-constrained, extremely limited dataset; system delivers response within the 2-second hard latency budget imposed by the biometric identification system.

• • Designed a framework-agnostic model compression system for neural network compression on low-power embedded hardware, with no dependency on specific model architectures.

• • Deployed anonymous on-device people counting systems in shopping malls and retail environments: video streams processed locally, zero data leaving the device, running without GPU on edge hardware.

• • Built CI/CD pipeline reducing deployment time from days to hours; migrated full infrastructure from AWS to GCP with infrastructure-as-code (Pulumi).

• • Implemented knowledge distillation pipelines on a custom CenterNet architecture, reducing model size while preserving accuracy; research on Transformer and Stable Diffusion compression.

• • NLP dataset creation for RL fine-tuning.

• • Created synthetic training data using Unreal Engine 4 to improve visual transfer learning for detection and re-identification.