Building a Real-World Adverse Media Pipeline
summary
Comply Advantage describes a deployed adverse‑media pipeline that processes 1 million articles a day, uses LLM APIs rather than hosted models, and balances a two‑stage prompt workflow against skyrocketing token costs and changing model versions. It documents governance like a model review board, human‑labeled golden datasets (with machine translation breaks in non‑Latin scripts), guardrails for hallucinations and PII, and operational lessons on rate limits, batching, and when fine‑tuning is actually worth the cost; the system is in production and working.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 0:54Thank you
- 1:55Who am I
- 3:45Comply Advantage
- 5:33What is Adverse Media
- 9:29What is not Adverse Media
- 10:23Why Gen AI
- 10:37Model proliferation
- 11:55One model to rule them
- 12:11Twostage process
- 15:18Upfront data scientists
- 15:45Personas
- 16:48Compliance Officer
- 19:08Magic Mike
- 21:15Start Simple
- 22:50Model Evolution
- 25:39Lessons Learned
- 27:24Model Performance
- 28:55The return of cautious CLA
- 29:17How do you evaluate
- 31:01CLA steps in wait
- 31:12Gold to cheese
- 33:02Machine translation
- 34:15Security
- 35:40Monitoring
- 36:23Conclusion
- 37:59We are hiring
related talks
33mMachine Learning to moderate ads
35mContent Creation in a Regulated Environment
16mProduction MLOps for Spam Detection: Real-Time Multi-Modal AI
42mOnline Safety Bill: How Global Platforms Use MLOps to Keep People Safe
11mDesigning Governed AI Pipelines for Healthcare Impact
19mMachine Learning for Privacy-Preserving Ads
30mML Automoderation System for LLM-Augmented Online Platforms
13mHuman-in-the-Loop MLOps: Boost Model Performance by 40%