Data classification with the Australian Government
summary
Australian Bureau of Statistics builds an AWS MLOps pipeline with SageMaker pipelines and Jupyter notebooks, migrating from on‑prem 600k/week training to cloud training of ~10M records in an hour. A hybrid approach uses hierarchical SVMs plus DistilBERT LLMs, backed by a confident‑learning UI for targeted data relabeling, producing ~9–10% point‑of‑capture coding, HSVM then LLM uplift to ~95% autocoding (leaving ~5% manual) — a ~20% clerical reduction versus 2021 — exposed via authenticated APIs including a 1 second response public service.
AI-written summary · about
chapters (8)
- 0:00Introduction to session
- 3:55Why ABS needed ML
- 9:18Meet the ABS team
- 12:10How autocoding works
- 17:30ML evolution at ABS
- 23:09HSVM vs large language models
- 27:42Government API service
- 29:53Key lessons learned
related talks
55mBuilding a secure MLOps pipeline, featuring PathAI
27mDetecting fraud faster with Amazon SageMaker, featuring Coinbase
35mA day in the life of a machine learning data scientist at JPMorgan Chase
31mBuild threat detection models using ML
29mSecure and compliant machine learning for regulated industries
21mSaaS Compliance Automation with a Python Stack
46mML for Anti-Money Laundering
42mBuilding a Robust Fraud Detection System