Peter Jung

- Improved quality of sales email writing assistant (currently in A/B test)

- Migrated in-house data pipeline into Databricks (running on Spark and Delta Tables), reducing the run-time of processing 60M emails from days to hours and halving the costs.
- Refactored 2 models to MLFlow and deployed on Databricks’ serverless endpoints.
- Fine-tuned and served GPT3 models.
- Wrote and debugged prompts for the best behaviour from pre-trained one/few shot models.

- Increased the accuracy of the multilingual sentiment analysis model written in PyTorch by 21%.
- Created a multi-modal (image and text input) model written in TensorFlow that reduces the workload of the other team by 84%.
- Delivered models that are running in production on millions of social media messages.
- Optimized existing models to have more than 50% faster inference speed and lower memory usage.
- Implemented a reverse image and video search engine with PyTorch and FAISS.
- Created an extreme text classification system with APIs for training and inference management, with an automatic training pipeline in Databricks.
- Assured that experiments are fully reproducible by properly using tools like MLflow, Git, DVC, Docker, and others.
- Started Python educational group.

- Delivered machine learning infrastructure based on Terraform (Terragrunt) infrastructure as code (IaC) on AWS.
- Configured AWS EKS (Kubernetes) with EC2 and Fargate workers.
- Deployed MLflow and Airflow to Kubernetes, including KEDA auto-scaling, XComs stored in S3, and workers configured for Fargate and EC2.
- Migrated manually managed EC2 instances to AWS ECS on Fargate.