POC to PROD: Hard Lessons from 200+ Enterprise GenAI Deployments
summary
Practices from 200+ deployments show concrete tradeoffs: prompt engineering often outperforms fine-tuning now; embeddings alone do not make a faceted query system; caching and UX hide latency but slow inference loses users. Concrete stack notes: Postgres + PGVector and OpenSearch for disk-backed vectors, Redis for fast (expensive) RAM search, Bedrock/SageMaker or Trn/Inferentia (claimed ~60% price/perf) for inferencing, batch Bedrock at ~50% off. Examples include multimodal video search using frame pooling and Titan v2 embeddings, an amplitude-spectrogram hack for sports highlights, and hospital deployments where chat beats voice.
AI-written summary · about
chapters (14)
- 0:00Who we are
- 1:07Customer success stories
- 2:13Multimodal search demo
- 3:58Video annotation insights
- 5:02Speaker background
- 6:47Enterprise AI challenges
- 8:10Building AI moats
- 9:36AWS AI infrastructure
- 12:02Key lessons learned
- 13:45Prompt engineering wins
- 15:10Evaluating AI systems
- 16:32Generative UI benefits
- 17:54Know your end user
- 18:35Optimizing economics
related talks
37mRound Table Discussion
18mAccelerate generative AI development with integrated data platforms
24mScalable GenAI in Production
27mPutting AI in Production
50mFrom Transactions to Intelligence: Postgres for Open AI & Analyti
40mRolling out Enterprise AI: Tools, Insights, & Team Empowerment
28mUse OpenSearch as Your Vector Search Database To Optimize GenAI Accuracy!
57mEnter The Brave New World of GenAI with Vector Search