Aymen Fkir

Architecting a scalable, end-to-end RAG pipeline for a French startup, focusing on robust data ingestion, optimized vector indexing, and retrieval consistency to deliver high-performance LLM-driven solutions.Architecting

- Architected an extensible Python synthetic data framework using OOP principles and inheritance to generate diverse, multi-domain logs while ensuring 100% seeded determinism for reproducible model validation.

- Co-led a data team of 10 collaborators, managing task delegation and technical workflows to deliver high-quality, schema-aligned datasets for the classification pipeline.

- Developed CLI tooling and Makefile integrations to streamline the data generation process, enabling users to produce structured datasets via simple command-line interfaces.

- Engineered robust CI and QA workflows by implementing unit tests for data quality and determinism, while debugging pipeline errors to ensure stable project builds.

• - Designed and implemented a scalable pipeline that processed over 20 GB of raw GPS data into detailed traffic flow maps across France.

• - Developed a traffic upscaling model using ensemble learning (CatBoost, XGBoost, LightGBM) to estimate real-world traffic from sparse probe data, achieving an R² of 0.83.

• - Enhanced DeepVolt's DLIA prediction model by integrating new traffic-based features (e.g., capture score, EV estimates, demand-supply ratios), improving average utilization prediction R² from 0.50 to 0.75.

• - Delivered a solution that saved the company over £40,000 by avoiding the purchase of processed mobility data.

• - Contributed to the launch of a new standalone traffic flow feature within the DeepVolt platform, supporting clients in data-driven EV infrastructure planning.