Zakaria Laabsi

Design realistic technical scenarios and problem sets in LLM inference optimization, GPU kernel design, and systems performance engineering. GPU inference optimization, custom CUDA kernel development, model-serving systems (vLLM, TensorRT-LLM, SGLang), quantization and optimizer design. High-throughput inference engine: continuous batching, paged KV cache, speculative decoding. Experience with distributed inference (tensor/expert parallelism, continuous batching, KV-cache optimization) Stack: vLLM, TensorRT-LLM, SGLang, CUDA, Triton, NCCL

Better agents with Context Graphs + Memory layer + Ontology.

Working on PayFit Copilot, AI assistant that improves payroll and HR management. PayFit Copilot has already generated over 230,000 conversations. It is used by 75% of SMEs managing their payroll on PayFit, and has reduced support tickets by 20%.

- Develop AI & LLM backend solutions. - Optimizing Retrieval Augmented Generation (RAG) with Hybrid Search - Collaborate with multidisciplinary team of research engineers and data scientists.

- Create a fine-tuned Mistral model via Direct Preference Optimization (DPO) + Layer-Selective Rank Reduction (LASER) - Deliver an operational MVP with a user interface, with a dedicated and secure AWS inference endpoint. - Cost reduction through autoscaling and scale-to-zero. Access logs and monitoring, implementation of custom metrics, endpoint management via API/CLI.

• Developed and showcased a TensorFlow and Keras-based poisoned Convolutional Neural Network model at a cybersecurity convention, emphasizing the necessity for robust AI models in mitigating security risks. • Implemented scalable cloud infrastructures on AWS (S3, Lambda, and RDS) with Terraform, achieving 40% better data retrieval efficiency and 30% lower operational costs in retail and supply chain sectors. • Streamlined API query automation using Python, boosting database efficiency by 65% through parallel, concurrent, and asynchronous programming techniques. • Serving Mistral 7B (LLM) on L4 GPUs running on Google Kubernetes Engine for a PoC mapped to business needs.

◆ Accompagnement des étudiants dans leurs trajectoires professionnelles - Partage d'astuces méthodologiques - Soutien émotionnel - Éclairage basé sur le parcours professionnel ◆ Ma démarche s'aligne sur des piliers clés - Organisation - Motivation - Transmission d’expériences

• Led the deployment of an Elasticsearch index, visualizing electronic signature KPIs for 39 regional banks, supporting 6000+ branches, 72000+ advisors, and 24+ million clients with intuitive, real-time national data insights. • Spearheaded the development and deployment of CI/CD DevOps pipelines using the ELK stack (Elasticsearch), Jenkins and Ansible, resulting in significant process optimization and efficiency gains. • Implemented a Kafka-based scalable architecture, customizing it to meet diverse business entity needs and enhancing data processing capabilities.

- Production and creation of video-educational content in mathematics (analysis, algebra, etc.) for 1st year undergraduate mathematics students. - Clarification work and guidance on points of difficulty often encountered. ⇒ This has enabled several students to tackle the 1st year mathematics syllabus with greater confidence, and to do well in their mid-term exams.