Zebra Technologies Corporation · NASDAQ:ZBRA
Zebra Technologies — deepsense.ai — Technical Presentation (September 2023)
22 slides from Zebra Technologies's technical presentation, Sep 14, 2023
Slide 1: Zebra DevCon 2023
Zebra GT) DevCon 2023 Connect | Learn | Build
Slide 2: Decoding the Future: The Potential of Large Language Models
Decoding the Future: The Potential of Large Language Models September 14, Madrid, Zebra DevCon 2023
Slide 3: We are AI experts helping companies implement AI-powered solutions
Deep technical expertise Generative AI Natural Language Processing We are AI experts Predictive Analytics Computer Vision helping companies implement MLOps and Data Engineering AI-powered solutions Extensive solution experience 200 commercial AI projects 120 world-class AI experts 10 years on the ma
Slide 4: Do LLMs think?
Do LLMs think?
Slide 5: The Evolution of LLMs
The Evolution of LLMs Predicting Keyword based Processing Predicting the the next letter systems forwards and next letter (Language automata) (Bag of words) backwards as if it (Generative LLMs) made sense (also RNN) Remembering simple Sequential processing Processing all at once phrases and patterns
Slide 6: The Impact of LLMs
The Impact of LLMs 97% 98% 30% of business owners believe of global executives agree Generative AI is expected to that ChatGPT will benefit that AI foundation models achieve a 30% share of the their businesses will enable connections overall AI market by 2025 across data types, revolutionizing where
Slide 7: LLM Use Cases
LLM Use Cases Question answering Answer questions about products or procedures, extract information from the knowledge base Summarize a long document e.g., news, meeting notes, Text summarization email, research paper Text generation Generate or improve the quality of text e.g., for emails or market
Slide 8: LLMs in the Zebra Ecosystem
LLMs in the Zebra Ecosystem New model of user interaction with devices Smart text interface Knowledge retrieval Text-to-SQL and data visualization Handheld Computers Recommendation Engine Forecasting, Planning Wearable Computers Workflow Assistance and Pricing Tablets Task Management and Communicati
Slide 9: Product Recommendation
Product Recommendation What are the parameters WHAT DO LLMs DO? of specific TV? The system uses LLMs to prepare How does it compare a summary of parameters with other TVs? for a specific TV Wearable Computer Wearable Computer New employee New employee Product Descriptions Center BENEFITS Improved cu
Slide 10: Workflow Assistance
Workflow Assistance The customer wants to return a product, but they don't have WHAT DO LLMs DO? the credit card with them, that they The system uses LLMs to search for used to make the original purchase. the solution in the Knowledge Base How to proceed? and provides a guideline Handheld Computer H
Slide 11: Task Management and Communication
Task Management and Communication What should I do today? WHAT DO LLMs DO? Give me more information The system uses LLMs to regarding the create a summary of tasks and to provide “receiving inventory” detailed instructions task on how to conduct a specific task Tablet Tablet Warehouse employee Wareh
Slide 12: Forecasting, Planning and Pricing
Forecasting, Planning and Pricing What is my forecasted volume WHAT DO LLMs DO? and revenue for next month The system uses LLMs to for my category? How many translate natural language into an SQL price changes do I need to query and pulls the required data make? Tablet Tablet Sales Manager Sales Man
Slide 13: Case Study Frontline Worker Co-Pilot
Challenge Managing frontline workers presents numerous challenges, such as coordinating varying skill levels, ensuring consistent performance, retaining valuable Case Study employees, and handling efficient knowledge transfers Frontline Worker Co-Pilot Solution GenAI-powered frontline workers co- pi
Slide 14: Frontline Worker Co-Pilot
Frontline Worker Co-Pilot Real-time support More efficient communication for employees between employees Answers based on Better customer service internal and external data and higher customer satisfaction
Slide 15: Application Architecture
Application Architecture Knowledgebase: Knowledgebase: Knowledgebase: Knowledgebase: Store Knowledge Knowledgebase: Store Knowledgebase: Store Knowledgebase: Store Knowledge Knowledgebase: Knowledgebase: Knowledgebase: Knowledge Store Franchise base: Store Store base: Franchise Franchise base: Store
Slide 16: Datasource Integration Details – Filesystem
Datasource Integration Details – Filesystem File Wiki File Storage Database Storage resources File Watcher New file File changed Event Queue Embedding Event Queue Embedding Row deletion Vector DB Vector DB Row creation
Slide 17: LLM Output Failsafes
LLM Output Failsafes Failsafe model: Metadata Failure limit reached User Augmentation: Output Valid Refined question Metadata validation question Increase Invalid temperature Log the failure
Slide 18: Too Many LLM Applications
Too Many LLM Applications 3rd party I want LLM LLMs are too quality expensive ?! I can’t run Small models multiple LLMs are inaccurate
Slide 19: One Model Can Be Multiple Models
One Model Can Be Multiple Models Output Base model Adapter Singular, Constant One per application B=0 Pretrained r Weights A = ℕ(0, δ^2) d Input
Slide 20: Application Architecture
Application Architecture Vector search: Store Vector Vector Vector Vector search: search: search: search: Store Store Store Vector Vector Vector search: search: search: Franchise Franchise Vector search: Store Franchise Zebra Augmentation: S S S Failsafe Metadata Fil W Fil W Fil W Q Q Q es iki es ik
Slide 21: Thank you
Thank you Contact: tomasz@deepsense.ai LinkedIn: linkedin.com/company/deepsense.ai
Slide 22: Zebra DevCon 2023
Zebra GT) DevCon 2023 Connect | Learn | Build
deepsense.ai — Technical Presentation (September 2023) on 6ix