AI vs. Karen: Real-Time Customer Sentiment Detection at Scale
summary
A live streaming pipeline is built that ingests customer messages via Apache Kafka and Apache Flink, calls models per message and writes sentiment into a real‑time OLAP store, visualized with Superset. Benchmarks show recent LLMs average >1s latency and high cost so they lag at scale; an open‑source encoder model is an order of magnitude faster but misses domain edge cases. A custom-trained modern BERT fine‑tuned with synthetic and labeled data raises accuracy and finally detects hostile users (demo: ~2,000 messages, 6 detections, 31 rebuttals).
AI-written summary · about
chapters (13)
- 0:00AI versus Karen
- 2:08AI real-time pipeline
- 3:54LLM for sentiment analysis
- 7:04Kafka and Flink architecture
- 11:19Live demo setup
- 18:29Real-time sentiment results
- 22:41Consumer lag detection
- 24:28LLM streaming limits
- 25:50Small language models
- 33:18Encoder-only transformers
- 34:39Fine-tuning with synthetic data
- 38:31Final demo with custom model
- 40:39Q&A
related talks
01h 06mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale
43mSentiment Analysis in Action: Building Your Real-time Pipeline
25mReal-Time Fraud Detection at Scale
56mSocial Listening, Data Extraction & Discovery
34mActionable Insights with Real-time Streaming Analytics of Customer Reviews
41mUsing AI to Build Fair and Equitable Workplaces
24mEngaging new and diverse audiences with ML
42mAI Math Models Help Hunting Cybersecurity Threads