Zoomer.ai · B2B SaaS, AI-powered data assistant
The first production machine learning work
Machine learning engineer, alternance
- NER mismatch rate
- Halved
Zoomer.ai was a B2B assistant for querying a company's own data. I co-led three augmented-data features while completing the machine learning engineering programme attached to the role.
Halving entity-recognition mismatches
Search analytics depended on recognizing customer-specific metrics, dimensions, and values inside a question. I benchmarked spaCy against other frameworks, retained it on production-performance grounds, and improved the model enough to halve the recognition mismatch rate.
Scheduled retraining kept the model current as customer data changed. Weights and Biases recorded model performance, and the auto-trained model was deployed with Docker on AWS.
Two supporting product features
Smart analysis used Lasso and tree-based feature selection to identify the main drivers behind a metric. A separate backend generated natural-language narratives from structured data and exposed the results through a FastAPI service.
Delivery practice
Unit and functional tests covered the modules and API behaviour, with code reviewed through GitLab merge requests. The company later closed after its funding was cut.
Stack
- Python
- spaCy
- transformers
- PyTorch
- scikit-learn
- FastAPI
- Docker
- AWS
- Weights and Biases
- GitLab