« All projects

AeroVision

A full data and machine learning stack, from ingestion to a monitored model in production, built solo in four weeks.

Predicting waves of flight delays per airport, starting from a live mock sensor API with failures, null fields and quality flags. No clean dataset, on purpose : the point was to build the whole chain the way it exists at work, not to train a model on a tidy file.

What was built

• An ingestion and transformation pipeline writing into a time series database, with raw, curated and operational layers kept separate.
• A secured product API reading the curated layer and writing only to its own user table.
• A dashboard for the served predictions and the health of the pipeline.
• Model training with experiment tracking and a model registry, plus drift detection on the incoming data.
• Continuous integration building and pushing container images, and monitoring on top of the whole thing.

Result

• Seven phases delivered in four weeks, each one deployed rather than left on a branch.
• A private forge with its own runner and container registry, replacing a heavier setup that ate the server memory.
• The stack was shut down on purpose once it had served its role, with database dumps and repository archives kept.

Stack