Boosting AI Innovation: Wie AI-Plattformen helfen, eure Use Cases auf die Straße zu bringen #ConCon24

Generative KI bietet vielfältige Einsatzmöglichkeiten im Geschäftsumfeld. Doch Herausforderungen wie Datenschutz und Compliance verlangsamen oft die Entwicklung und Markteinführung von KI-basierten Produkten. Die Beteiligung verschiedener Stakeholder führt zusätzlich zu Verzögerungen.

AI-Plattformen eilen zu Hilfe! Wir beleuchten in diesem Vortrag, wie eine Corporate-AI-Plattform aussehen kann und welche Bausteine und Services sie benötigt.

Wir diskutieren, wie eine firmenweite Plattformstrategie nicht nur die technische Umsetzung vereinfacht, sondern auch ein Ökosystem für Innovation schafft, die Zusammenarbeit fördert, Wiederverwendbarkeit erhöht und letztendlich die Time to Market verkürzt.

#ConCon24 @qaware #AI #PlatformEngineering

Transcript

1. qaware.de Boosting AI Innovation Wie AI-Plattformen helfen, eure Use Cases auf die Straße zu bringen Mario-Leander Reimer mario-leander.reimer@qaware.de @LeanderReimer #CloudNativeNerd @qaware
2. 2 Mario-Leander Reimer Managing Director | CTO @LeanderReimer #cloudnativenerd #qaware #gernperDude
3. Platform engineering is the discipline of designing and building toolchains and workflows that enable self-service capabilities for software engineering organizations in the cloud-native era. Platform engineers provide an integrated product most often referred to as an “Internal Developer Platform” covering the operational necessities of the entire lifecycle of an application. https://platformengineering.org/blog/what-is-platform-engineering
4. “Too much cognitive load will become a bottleneck for fast flow and high productivity for many teams.” ▪ Intrinsic Cognitive Load Relates to fundamental aspects and knowledge in the problem space (e.g. languages, APIs, frameworks) ▪ Extraneous Cognitive Load Relates to the environment (e.g. console command, deployment, configuration) ▪ Germane Cognitive Load Relates to specific aspects of the business domain (aka. „value added“ thinking)
5. An IDP and your platform engineers are key enablers for high productivity of the stream-aligned DevOps teams. QAware | 5 ▪ Responsible to build and operation a platform to enable and support the teams in their day to day development work. ▪ The platform aims to hide the inherent complexity to reduce the cognitive load for the other teams. – Standardization (Compliance, Security, …) – Developer Self-Service ▪ Fully automated software delivery is the goal! https://hennyportman.wordpress.com/2020/05/25/review-team-topologies/
6. AI platform engineering is the discipline of designing and building toolchains and workflows to provide self-service capabilities for data and AI driven organizations. Business experts, data engineers as well as software engineers work together in an integrated platform from now on referred to as an “Enterprise AI Platform” covering the operational necessities of the entire lifecycle of AI use cases. © 2024, M.-Leander Reimer
7. Endless Possibilities and Use Cases Chatbots, CWYD, Content Creation
8. The most common uses for GenAI tools are in marketing, sales, product development and service operations. Source: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2023-generative-ais-breakout-year
9. Use Case: Customer Support Support Ressourcen Recherche Support Flow Customer Data Images and Icons were generated with the assistance of AI RAG Automation Intent Recognition Text2Speech Speech2Text Anomaly Detection Similarity Matching AI Assistant Call Chat Multi Agent Workflow
10. RAG in a Nutshell. Index, e.g. Vector DB Indexing (Chunking & Embedding) Documents Ingestion Phase Query Encoding Retrieval Phase Context Prompt LLM with world knowhow Response
11. From input to embedding: this is how a high-performance semantic search using vector databases works. Embedding Model Images were generated with the assistance of AI { 23.567, 45.899, 76.345, …}
12. Chatbots and AI Assistants The more specific the use case, the more complex it becomes. ChatGPT or comparable with world knowhow ChatGPT with organisational context knowledge Specialized AI Assistent ▪ Retrieval Augment Generation ▪ Transfer Learning ▪ Specially trained model ▪ Process automation Complexity Benefit ▪ Easy to use and cost efficient ▪ Requires guidelines on data protection and compliance
13. The 80% Fallacy Juan Pablo Bottaro, LinkedIn Engineering Blog
14. Why do we need an AI platform?
15. Key challenges: models, tools and skills. Source: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2023-generative-ais-breakout-year
16. Each stakeholder involved has a different expertise and thus a different focus. Domain Expert Software Engineers and Architects This image was generated with the assistance of AI Data Scientists, AI Experts Platform Engineers
17. It all starts with an understanding of data and use cases. This is crucial for a structured start in any AI project. Business Understanding Data Understanding Data Preparation Modelling Evaluation Deployment
18. "But we are already doing this!" Really? MLOps only covers part of the tasks related to GenAI. Source: https://neptune.ai/blog/mlops
19. Our proposal for an AI platform architecture
20. Integration & Delivery Plane Service Plane Platform Plane Resource Plane Quality Plane Compliance Plane Foundation Foundation Interfaces Domain Services Domain Services Domain Services
21. Compliance Plane Integration & Delivery Plane Service Plane Data Plane Platform Plane Observability Operability Resource Plane User Serving Plane Access Plane / APIs Orchestration Plane Data Modelling Plane Model Plane Compute Data Integration Security Delivery FinOps Quality Plane
22. Quality Plane Integration & Delivery Plane Service Plane Access Plane/APIs User Serving Plane Orchestration Plane Data Modelling Pl. Data Plane Model Plane Compliance Plane Platform Plane Observability: Monito- ring, Logging, Tracing Security: Secrets, IAM Encryption, Certs, … Scale, Backups, Recovery, … Delivery: CI/DC, Registry Pipelines, Orchestrator, … FinOps Resource Plane Compute: CPU and GPU Data: Vector DBasS, other Storage, … Integration: Self-hosted LLMs Public LLMs Managed AI Services
23. Quality Plane Integration & Delivery Plane Service Plane Access Plane/APIs User Serving Plane „Convenience UIs“, Self Service, RAG per Drag and Drop, … (a) LLM, Embedding, (b) RAG, Chatbot, … (c) Data Access, … Orchestration Plane Data Modelling Pl. Playground Prompt Engineering Konfiguration Runtime, Instantiation, Orchestration, Scaling, Configuration Data Plane Model Plane Compliance Plane Platform Plane Observability: Monito- ring, Logging, Tracing Security: Secrets, IAM Encryption, Certs, … Scale, Backups, Recovery, … Delivery: CI/DC, Registry Pipelines, Orchestrator, … FinOps Resource Plane Compute: CPU and GPU Data: Vector DBasS, other Storage, … Integration: Self-hosted LLMs Public LLMs Managed AI Services
24. Quality Plane Integration & Delivery Plane Service Plane Access Plane/APIs User Serving Plane Technical and Business Metrics like Accuracy, Harmfulness, … Test Automation for LLMs „Convenience UIs“, Self Service, RAG per Drag and Drop, … (a) LLM, Embedding, (b) RAG, Chatbot, … (c) Data Access, … Orchestration Plane Data Modelling Pl. Playground Prompt Engineering Konfiguration Runtime, Instantiation, Orchestration, Scaling, Configuration Data Plane Ingestion Pipelines Data Versioning Embeddings & Vectorization Model Plane MLOps: Model Registry Model Management Experiment Tracking Model Serving Compliance Plane Tonality, Bias Security, Data Protection Platform Plane Observability: Monito- ring, Logging, Tracing Security: Secrets, IAM Encryption, Certs, … Scale, Backups, Recovery, … Delivery: CI/DC, Registry Pipelines, Orchestrator, … FinOps Resource Plane Compute: CPU and GPU Data: Vector DBasS, other Storage, … Integration: Self-hosted LLMs Public LLMs Managed AI Services
25. From concept to realisation: possible variants
26. All roads lead to Rome. Depending on the context, one or other variant makes sense. Buy an AI platform solution Combination of cloud provider building blocks Custom platform with open source components
27. Azure AI Studio (Preview) Azure AI Content Safety Quality Plane Integration & Delivery Plane Service Plane Azure API Management Access Plane Azure AI Studio (Preview) User Serving Plane Azure AI Studio (Preview) Semantic Kernel Orchestration Plane Azure AI Document Intelligence Data Modelling Pl. Azure AI Search with Indexers, Indices incl. Vector DBs. OneLake, Fabric Data Plane Azure OpenAI Azure Machine Learning Model Plane Azure AI Content Safety Compliance Plane Platform Plane Observability Security Scale, Backups, Recovery, … Delivery FinOps Resource Plane Compute Data Azure OpenAI Azure AI Language Speech Service Azure AI Translator Integration Overview on Azure AI Services: https://learn.microsoft.com/en-us/azure/ai-services/what-are-ai-services
28. Azure AI Studio (Preview) Azure AI Content Safety Quality Plane Integration & Delivery Plane Service Plane Azure API Management Access Plane Azure AI Studio (Preview) User Serving Plane Azure AI Studio (Preview) Semantic Kernel Orchestration Plane Azure AI Document Intelligence Data Modelling Pl. Azure AI Search with Indexers, Indices incl. Vector DBs. OneLake, Fabric Data Plane Azure OpenAI Azure Machine Learning Model Plane Azure AI Content Safety Compliance Plane Platform Plane Observability Security Scale, Backups, Recovery, … Delivery FinOps Resource Plane Compute Data Azure OpenAI Azure AI Language Speech Service Azure AI Translator Integration Overview on Azure AI Services: https://learn.microsoft.com/en-us/azure/ai-services/what-are-ai-services Just give it a try. Or ask Azure experts.
29. mlflow, Evidently AI, RAGAS (for RAG), DeepEval (for LLM) Quality Plane Integration & Delivery Plane Service Plane API Gateways Access Plane Build your own User Serving Plane Kubeflow Orchestration Plane Jupyter Kubeflow Data Modelling Pl. Weaviate, neo4J, … Custom Pipelines Data Plane mlflow (Registry) BentoML (Serving) Kubeflow (Serving) Model Plane Build your own Compliance Plane Platform Plane Observability Security Scale, Backups, Recovery, … Delivery FinOps Resource Plane Compute Data LLMs: Llama, Mistral, … mlflow BentoML Integration
31. mlflow, Evidently AI, RAGAS (for RAG), DeepEval (for LLM) Quality Plane Integration & Delivery Plane Service Plane API Gateways Access Plane Build your own User Serving Plane Kubeflow Orchestration Plane Jupyter Kubeflow Data Modelling Pl. Weaviate, neo4J, … Custom Pipelines Data Plane mlflow (Registry) BentoML (Serving) Kubeflow (Serving) Model Plane Build your own Compliance Plane Platform Plane Observability Security Scale, Backups, Recovery, … Delivery FinOps Resource Plane Compute Data LLMs: Llama, Mistral, … mlflow BentoML Integration Use at your own risk! Or ask an AI platform expert.
32. Which one is right for me?
33. Start lean and agile! Tailored to the domain and problem instead of ‘One size fits all’. Use Case Identification Business Understanding Skill, Resource & Requirements Analysis Building Block Mapping & Prioritization Implementation Evaluation Commoditization
34. QAware GmbH | Aschauer Straße 30 | 81549 München | GF: Dr. Josef Adersberger, Michael Stehnken, Michael Rohleder, Mario-Leander Reimer Niederlassungen in München, Mainz, Rosenheim, Darmstadt | +49 89 232315-0 | info@qaware.de Thank you!