Data scientist turned ML engineer - From notebooks to production

General
1

Hello TianPan community! Rachel here, data scientist who evolved into ML engineering.

Started analyzing spreadsheets, now deploying models at scale!

My evolution:

  • 2018: Excel pivot tables (simpler times)
  • 2019: Discovered Python and pandas
  • 2020: Deep learning obsession begins
  • 2021: First model in production (crashed immediately)
  • 2022: Learned MLOps the hard way
  • 2024: Building ML platforms

Current tech stack:

  • Modeling: PyTorch, JAX, HuggingFace
  • MLOps: Kubeflow, MLflow, Weights & Biases
  • Data: Spark, Dask, Ray
  • Deployment: FastAPI, BentoML, Triton
  • Monitoring: Evidently AI, WhyLabs

Hard-learned lessons:

  • Jupyter notebooks are not production code
  • Data drift kills models silently
  • Feature stores are worth the complexity
  • Model performance != business value
  • Explainability matters more than accuracy

Current projects:

  • Real-time recommendation system (1M+ requests/sec)
  • LLM fine-tuning pipeline
  • Edge ML deployment framework
  • Federated learning experiments

Building an open-source tool for model monitoring. Beta testers welcome!

Who else is doing ML in production? What are your biggest pain points?

2

Welcome Rachel! Full-stack dev here working with ML teams. Your point about notebooks not being production code hits hard - we learned that lesson painfully. How do you handle the handoff from data scientists to engineers?

3

Hi Rachel! Security engineer here. Model monitoring is crucial for detecting adversarial attacks. Your federated learning work sounds fascinating - how do you handle privacy-preserving training at scale?