Tahar Souici

Context: Aviv Group is a leading European digital real estate company, operating property platforms across multiple countries. Within the Data team, I work on the design and optimization of geographical referential data systems that are critical to powering real estate search platforms across Europe. I also contribute to R&D initiatives exploring multimodal AI for innovative property search experiences.

▷ Geographical Data Platform

- Maintenance and optimization of geographical referential data pipelines supporting real estate platforms across multiple European countries.

- Design, development, and operation of large-scale APIs ensuring high availability, performance, and data consistency across geographies.

- Definition and implementation of data system architecture, enabling robust and scalable multi-country geographical data integration.

- Led stack optimization efforts across infrastructure, data pipelines, and API services.

- Mentoring of junior engineers, systematic code reviews, and promotion of engineering excellence through documentation and standardization.

▷ R&D — Multimodal RAG (Text-to-Image Search)

- Designed and built a multimodal RAG proof of concept enabling users to search for properties by describing their desired apartment style in natural language (layout, decoration,ambiance...).

- The system converts textual descriptions into embeddings and performs similarity search against listing images from the platform’s catalogue to find visually matching properties.

- Built an end-to-end pipeline combining LLM, multimodal embeddings, and vector search over the property images database.

- This POC paves the way for a groundbreaking property search experience, going beyond traditional filters (price, area, location) to offer visual intent-based search.

Technical Environment: Python, DBT, REST APIs, Git, CI/CD, Microservices Architecture, LLM, RAG

Context: Decathlon, the world’s largest sporting goods retailer, undertook the construction of a self-service Data Factory for ingestion to industrialize and democratize data access at group scale.

Within a cross-functional team (1 PO, 1 Tech Lead, 1 DevOps, 1 Fullstack, 3 Data Engineers), I played a central role in the design, development, and evolution of this ingestion platform handling massive volumes of heterogeneous data.

▷Self-Service Ingestion Platform

- Designed and developed a self-service data ingestion platform, enabling business teams to orchestrate their own ingestion workflows autonomously.

- Ingestion of over 1,000 heterogeneous data flows.

- Developed data pipelines in Scala and Spark following a medallion architecture on AWS.

- Integrated Databricks Autoloader for high-volume use cases, ensuring performant and reliable incremental ingestion.

- Developed a feature enabling the integration of Airbyte connectors into the platform, significantly expanding the catalogue of supported data sources.

▷Cross-Cloud Ingestion Agent

- Designed and developed a cross-cloud ingestion agent enabling data collection from Alibaba, GCP, and Azure into AWS.

- Agent built with Scala and ZIO (functional programming), ensuring robustness, scalability, and performance: ingestion of multiple GB of data with zero failure.

▷Observability & Data Lineage

- Established a platform-wide data observability strategy, enabling proactive anomaly detection and trust in ingested data quality using Great Expectations to automatically validate data quality, and OpelLineage for end-to-end data lineage tracking.

Business Impact: Platform used daily by dozens of Decathlon teams worldwide. Drastic reduction in time-to-data for business teams through self-service. Unification of multi-cloud data flows at international scale. Full data observability and lineage enabling confident, governance-compliant data consumption.

The SNCF AIFluence R&D project aims to predict passenger crowd levels in French railway stations to optimize traveler flow management and improve the customer experience. I was tasked with industrializing Machine Learning models developed by Data Scientists and building the MLOps infrastructure required for their production deployment.

- Built a complete MLOps stack for the industrialization of crowd prediction ML models.

- Code refactoring: transformed exploratory notebooks into production-grade, structured, and maintainable projects.

- Established software engineering best practices: modular code organization, Pull Request workflows, unit and integration testing, technical documentation.

- Deployed and configured Kubeflow for model training, versioning, and production monitoring.

- Set up CI/CD pipelines (GitLab CI) for automated deployments and experiment reproducibility.

- Infrastructure management: provisioning and maintenance of environments via Terraform and Kubernetes on AWS.

Business Impact: R\&D project that successfully delivered the industrialization of crowd prediction models. The MLOps stack fundamentally transformed the Data Scientists' working environment: transitioning from exploratory notebooks to an industrialized, reproducible, and deployable product. Teams now have a structured framework to train, version, and monitor their models autonomously, dramatically reducing the cycle from experimentation to production.

Python, Terraform, Kubernetes, Kubeflow, GitLab CI, AWS, Docker, CI/CD, MLOps, Infrastructure as Code