This is not rocket (data) science
summary
Team ships 43 apps in three years and finds none use AI; practical builds include an on‑device sentiment colorizer, a translation button backed by iOS 12 language recognizer, a search bar powered by 300‑dimensional natural language embeddings, and a simple intent classifier trained from email data. Uses BERT SQuAD (~220MB) for QA, Firebase smart‑reply pod, and TensorFlow Lite .tflite inspection; models work but have limits: English bias, model size, and modest precision — consensus ensemble hits about 88% accuracy.
AI-written summary · about
chapters (13)
- 0:00Start
- 4:09Voice interfaces
- 5:54Translation
- 15:33Visualize NLEmbedding
- 16:49NLEmbedding : nuances
- 18:22NLEmbedding: lexical filter
- 20:38Firebase smart-replies
- 21:26How does it work?
- 21:45Hobbes... Leviathan ?
- 22:56Sensitive model...
- 25:55Requests-for-action
- 28:43Improve accuracy
- 29:20Mendeleev Boring Al Classification
related talks
01h 00mData, visualization, and designing AI
47mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale
53mVisualizing the Output of Your Search AI
55mKeynote – Joseph Sirosh
14mSolving App discovery through smarter user data analysis
13mThe 2025 AI Engineering Report
26mUber’s Intelligent Insights Assistant
55mSocial Listening for Richer Data, Unstructured Analytics