Every January a new “AI engineer stack” list drops. Most have a launch sponsor. Here is what we actually use at LIACC, unsponsored.

Daily drivers

  • uv for Python package management. Fast, boring, correct.
  • Ruff for linting and formatting.
  • Polars for data wrangling; Pandas only when we're sharing notebooks with a statistician.
  • DuckDB for analytical SQL over files.
  • HuggingFace Hub + vLLM for inference.

Training + evaluation

  • Axolotl for SFT/DPO on open weights.
  • LM-Eval-Harness for baseline benchmarks.
  • Our own small task-specific eval suites. No substitute.
  • Weights & Biases for experiment tracking.

Shipping

  • Modal or Runpod for burst GPU.
  • Triton for kernel work.
  • FastAPI + Pydantic for the gateway.
  • Grafana for dashboards we can debug in an incident.

Two things we stopped using

  • Langchain, as a required dependency. We still import it occasionally.
  • Notebooks for anything that needs to ship. Scripts with types, please.