Why Risk Management Requires Taking Risks: A Practical Guide...
summary
Eighteen months of production work on live game streaming (150+ services, 25k+ servers, ~100k GPUs) produces concrete patterns: deploy a RAG server for grounded answers, host on-prem models with NVIDIA NIMS to satisfy security, expose read-only tools to surface alerts and SLO time series, and combine a Bayesian network to rank likely causes. Built pipelines generate candidate remediation commands, run multi-step filters and human approval, and pilot automatic rollout restarts that reduce time-to-resolution while containing hallucination and alert-fatigue risks.
AI-written summary · about
chapters (24)
- 0:00GeForce Now and SR team
- 1:23Managing 150+ global services
- 2:27AI in incident response
- 3:30Reducing misescalations with AI
- 4:53Streamlining reports and postmortems
- 5:58Handling non-SR questions with AI
- 7:01Filling skill gaps with AI copilots
- 7:43Why AI adoption feels risky
- 8:47Overcoming security pushback
- 9:30Executive support for AI projects
- 10:12Starting with AI summarization
- 11:37Synthesizing real-time info with AI
- 12:41AI evaluation of system problems
- 13:23AI action on live systems
- 14:46Practical AI implementations by Arvin
- 17:37Building RAG server for docs
- 20:06Connecting AI to monitoring signals
- 21:56Detecting anomalies with forecasting
- 23:39Generating incident reports
- 25:24Weekly reliability digests
- 26:29Plotting SLO data on-the-fly
- 27:11Bayesian networks for root cause
- 30:39AI resolutions with human oversight
- 33:10Conclusion: AI for risk management
related talks
07mIf Development is a Game, How Do I Win?
27mImplementing Chaos Engr. in a Risk Averse Culture
27mAI-Powered Reliability: Stop Fighting Fires, Prevent Them
47mPersistent SRE Antipatterns: Pitfalls on the Road to Creating a Successful
11mManaging business risks together
52mPersistent SRE Antipatterns: Pitfalls On the...
25mLLM Agents for Site Reliability
30mNavigating the New AI Security Landscape