Rabobank Data Expo 2024

Transcript

1. Structuring Wholesale Sustainability Information using Generative AI Gen AI Sustainability Database
2. 8000 hours What if I tell you that… Retrieving sustainability information
3. What if I tell you that… 3 > 90% Accuracy
4. 70,200 Reduced time by 40% Scalable Faster Cheaper Benefits Easily Scale up Why?
5. How did it start? 5 2022, I was working on a cool project. BERT
6. 2023 6 The bank started experimenting with Generative AI
7. 2024 - Sustainability information 7 Waste (food waste, reduce, reuse, and/or recycle waste) Water (Water usage, water productions, water treatment, water target, challenges and opportunities etc) GHG Emissions (Scope 1, 2, 3, emission reduction target, net zero target, ) Climate (renewable energy share, renewable energy target) Biodiversity (biodiversity target, assessment of biodiversity challenges, deforestation, sustainable sourcing of materials etc) Current focus Future topics
8. How do you work with GenAI in 2024? 8 import openai Prompt = “ You are a helpful assistant…” import o tab tab *waiting* tab tab
9. Generative AI is so amazing…. 9
10. Expectation 10 Happy stakeholders
11. That did not happen 11
12. Well… 12
13. We have very bad outcome at first… 13 30% accuracy
14. Then… ~70% accuracy Use better and larger model Teach the model sustainability concepts Few-shot learning
15. 15 Issue - Tables without clear borders
16. Issue - Complicated infographic 16
17. Then… 17 Processing documents with ADOBE
18. Then… 18 Retrieval Augmented Generation ( aka RAG)
19. 19 Retrieval Augmented Generation (RAG) Marginal maximal Relevance
20. Then… 20 Chain-of-thought/reasoning
21. Chain of thought reasoning 21
22. Then… 22 AI Self-Critique
23. Self-critics 23
24. Then… 24 82% accuracy Self-Critique Processing documents with ADOBE Chain-of-thought/reasoning Retrieval Augmented Generation
25. And then… 25 Prototype, collect feedback, evaluate and iterate
26. And the accuracy now… 26 Green house gas emission metrics
27. 27 Renewable Energy metrics And the accuracy now…
28. Demo User Interface 28
29. Take-aways 70,200 • It is easy to build cool demo • It takes lots of effort to build a final product • AI is a new thing, but have huge potential • Experiment, see what works for you • Have fun, learn and share the knowledge
31. Evaluation 70,200 • Compare AI generated output with ground truth • Consistency of model output • Human Validation • AI as a judge (Self-reflection)
32. Consistency in model output 32