《使用基础模型快速扩展应用的 AI ML并将其应用于现代 AI ML 用例.pdf》由会员分享，可在线阅读，更多相关《使用基础模型快速扩展应用的 AI ML并将其应用于现代 AI ML 用例.pdf（32页珍藏版）》请在三个皮匠报告上搜索。

1、Start from scratch valueLeveraging Foundational Models to Rapidly Scale Applied AI OutcomesNick King,Founder/CEO Data K35%of projects miss expectations.90%of S&P 500 companies now publish ESG reports in some form.McKinseyCustomer expectationsInvestor demandRegulatory requirementsCompetitive position

2、ingCorporate reputationOrganizations face mounting pressure to document and improve ESG performance$20BEvery year,Fortune 500 companies devote approximately$20 billion to their CSR efforts.ForbesTwo thirds of an average companys ESG footprint lies with suppliers.For some organizations with tens of t

3、housands of suppliers,its paramount to accurately capture their ESG footprint.Relying on manual processes,questionnaires and investigations is not only highly resource-intensive,but its also extraordinarily inaccurate.Dear Supplier,how ethical are you?A SMARTER ALTERNATIVEAutomate supplier ESG compl

4、iance at scale using LLM and Applied AI/MLHow does Applied AI accelerate this?Risk Identification and AssessmentAI can analyze vast amounts of data to identify potential ESG risks in the supply chain.It can assess the likelihood and potential impact of these risks,helping businesses prioritize their

5、 risk management efforts.Monitoring and ReportingAI can continuously monitor various data sources for new information that might indicate changes in ESG risk levels.It can also automate the generation of detailed ESG compliance reports,saving time and reducing the risk of human error.Predictive Anal

6、yticsAI can use historical data to predict future ESG risks.This can give businesses more time to develop and implement risk mitigation strategies.Decision SupportLarge language models can provide decision-makers with insights and recommendations based on their analysis of ESG data.This can support

7、more informed and effective decision-making around ESG risk management.Stakeholder EngagementAI can help businesses engage with stakeholders(including investors,customers,and local communities)on ESG issues.This can include identifying key concerns and questions from stakeholders,developing response

8、s,and tracking stakeholder sentiment over time.12345Why do LLMs and Applied AI work so well?Large volumes of unstructured dataExtraction and VectorDBESG baselines can be inferred within industriesLDA and ML scoringOther sources(news,invoices)can uncover additional detailed insightsExtraction,VectorD

9、B,LLM for summarizationBringing it all togetherLLM for summarization,ML models for heat mapping1234CSRAzureBlobFeature StoreESG Tabular DataAzureFlinkDK Extract1Tabular to LLMApproachApproachADAText20222021Ingestion PhaseTextSplit text into chunksEmbedding model0.1,0.2,0.3202220210.1,0.2,0.3Vector S

10、torePineconeSimilarity Search and RetrievalESG Gap AnalysisBotInvestment AnalyzerUser Front EndTopic Models BuildCustom ESG ScorePredictions and InsightsPDFsAzureDBAzureDBAzureDB24568921PDFsAzure OpenAI(Private)Additional LLM models as requiredESG App ArchitectureLeveraging Databricks,LLM

11、,and Applied AIDatabricks DollyGood for summarizationFAISSVectorDBAzure FeathrFeature StoreDatabricks ML,pipeline 1_DAIS_Title_SlideHow does it work?Summarize complex CSR reports into a specific ESG related themesPhase 1Phase 2Phase 3LDAExtract news data and align to company and categorizationMany t

12、echniquesPhase 1Phase 2Phase 3DemoBringing it togetherGoing from models to outcomesGoal is to then converge all data into one VectorDB,and create agents to query both the models,and the LLM to summarize data against companies,categories,and inferences.Considerations:What private data needs to be kep

13、t isolated?How do you build trust scores news sources?How do you prevent hallucinations?123Leverage Dolly 2.0 or other Open Source model(good for summarization tasks)Match quality and quantity scores based on summarization and inference modelsConverge extraction into both a feature store,and embeddi

14、ngs into Vector store1_DAIS_Title_SlidePreventing Hullication and building trustCreate anti-hallucination checks with model inferencePhase 1Phase 2Phase 3Create a control loop with the data and the resultsBuild both tabular data and embeddingsCompute features from data to validate outputs from agent

15、sTrain models to infer outcomes to support further validationCSRAzureBlobFeature StoreESG Tabular DataAzureFlinkDK Extract1ADAText20222021Ingestion PhaseTextSplit text into chunksEmbedding modelPDFs2456PDFsSet temperature&manage concept driftPhase 1Phase 2Phase 3Zero CreativityTemperate=0.0 alwaysEn

16、sure that prompts and responses are baselined(concept drift)-promptfoo etc Validate outcomes from agent with inferred values to ensure result within correct rangeConfigure retrievalPhase 1Phase 2Phase 3Converge ML w Search to validate outcomesThis approach reduces risk in hallucinations,but also pro

17、vides referenceable metrics for use with the businessTip:One step at a time,ensure that you can predict outcomes and have high confidence before moving to the next1_DAIS_Title_SlideAgent Configuration and preventing sensitive data leakageAgent ConfigurationPhase 1Phase 2Phase 3Create multiple agents

18、 that act as expertsBy creating specific expert agents place controls and boundaries around responses.Work with the concept of a query router that specifically calls particular models based on the performance and data sensitivity requirements.Building on the Applied approachBecomes possible to creat

19、e high trust,detailed insights that can be rapidly summarized.Recommendations become their own agent of excellence.Chat is reduced down to work only on available data reducing risk of hullicationIdentifying ESG gaps:Are companies walking the talk?Phase 1Phase 2Phase 3Best practices summaryBuild to s

20、caleStart smallUse ML/AI+LLMEmbeddings create efficiency when done rightUnderstand models and their capabilities1_DAIS_Title_SlideWhy Data Kinetic?Data Kinetic ExtractDK Extract uses a combination of Natural Language Processing(NLP)and Generative AI to extract key information from invoices.This incl

21、udes details such as the invoice ID,total amount,and date.This automated extraction is efficient and accurate,saving time and reducing the risk of human error.Designed to handle a variety of document types,including images embedded inside PDFs,regular but unstructured PDF templates,and regular PDFs

22、with structured tables.Video,Audio,and any document asset can be extracted.This versatility makes it a useful tool for a wide range of applications.DK Extract integrates multiple AI techniques to handle different types of invoices.For instance,it uses a combination of computer vision,generative AI,a

23、nd NLP for images embedded in PDFs.For regular PDFs,whether structured or unstructured,it uses a combination of an expert system,generative AI,and NLP.This multi-modal AI approach allows DK Extract to adapt to the specific requirements of each invoice type,improving its effectiveness and accuracy.Ef

24、ficient Information ExtractionHandling of Various Invoice TypesIntegration of Multiple AI TechniquesMulti-modal extraction Foundational ModelPrivate LLM at scalePrivate LLM deployments and outcomes in your environmentComplex use case blueprintsAccelerating the path to valueKeyHow does it work?1/Sele

25、ct your industry outcomeInsuranceFinancial ServicesOil&GasHealthcareForecast patient volumePredict equipment failurePredict patient staysFraud detectionRisk assessmentPolicy recommendationsPortfolio optimizationDeal identificationPredict product exposurePredict emissions3/Deploy in your environment

26、of choiceDeclarative Models Foundational Models2/Rapidly build AI applications using proven,repeatable components1/Customers start by selecting a high impact use case from the AI Application Catalog that best meets their needs.2/Data Kinetic then rapidly customizes the AI application using a mix of

27、custom and third-party components.A proprietary data technique transforms and augments all data into a ready-to-query format via the Engine.3/Data Kinetic will then deploy on top of Snowflake,Databricks,Palantir,or a customers existing ML platform.Outcomes are maintained remotely from Data Kinetics solution fabric.Data Extraction Engine ExtractNovelIn ProgressA new approach to creating Applied AIDelivering AI outcomes with meaningful