Creating a scalable (...) with Elixir, Python & Tensorflow
summary
Built a multi-tenant intent service called Q&A ninja that uses Elixir to orchestrate pools of Python workers and a shared sentence-BERT embedding service so a 340M‑param model is not duplicated per bot. Each bot trains a small TensorFlow DNN classifier (77 intents in the demo) in about 10 minutes and achieves ≈91% precision. Python worker processes cost about 300 MB each and hold multiple models; initial high false‑positives are trimmed by adding negative utterances, and the embeddings prove unintentionally multilingual.
AI-written summary · about
chapters (13)
- 0:00Intro
- 1:41Context
- 9:18Intent classification
- 12:46Sentence feature vector
- 18:22Base training set
- 20:38Rasa results
- 21:47Scaling up
- 24:08Demo time
- 31:49Python: trainer script
- 34:18Python: classifier script
- 35:36OnA classifier outcome
- 39:26Noteworthy Elixir libraries
- 40:47Future work..
related talks
39mFrom data science to scalable NLU and vision cloud service.
33mBuilding the next-generation Conversational AI with Python and Deep Learning
23mUnveiling the Journey of NLP
31mMachine Learning-Enabled Chatbots
38mNatural Language Understanding at Scale
40mBuilding highload ML powered service
42mDevOps & Microservices: “Delivering Machine Learning models early and often.”
28mByron Galbraith, Chief Data Scientist, Talla at MLconf Seattle 2017