This is the second in a series of three posts covering DataSapien edge Architecture. This post covers the DataSapien Personal Intelligence capabilities. DataSapien offers Personal Marketing Technology to brands. Its edge architecture is focused on providing better customer outcomes. It is comprised of a bundle of components that are embedded inside existing customer apps, as a Software Development Kit, or SDK.
With the DataSapien edge architecture SDK, personal data held on the edge (on the device) is also processed on the customer’s edge device, utilising the DataSapien Personal Intelligence stack. This is similar to the new Apple Intelligence approach to moving intelligence out to devices and with DataSapien Brands can tap into similar technology. In a Brand and Customer context; DataSapien Personal Intelligence enables Brands to create personal customer journeys that prioritise and serve the jobs that customers hire the brand to get done for them. Brands optimise the delivery of customer outcomes that align with their Brand Promise.
The DataSapien Personal Intelligence architecture comprises three layers:
2.1 Deterministic Intelligence
A sophisticated rules engine which runs deterministic ‘if-this-then-that’ actions locally on the data in the DataVault. The rules are defined in The Orchestrator (see below). This creates logical decision flows, surfaces insights, generates signals and conducts customer journeys on the customer’s edge device. This is ideal for activities such as the local generation of Next Best Actions (NBA) for the customer.
Deterministic logic can also be used to process rating scales in the app. These might include psychometric insights or customer segmentations, for example, framed as “your shopper personality type”.
Privacy Enhancing Technologies (PETs) such as Homomorphic Encryption will extend these capabilities to interrogate and process data while it remains encrypted, utilising the key strengths of the edge architecture approach.
2.2 Probabilistic Intelligence
Machine Learning models can be deployed into the local SDK to process data and make probabilistic inferences and suggestions on the edge. These models are trained centrally and are deployed locally on-device, using the Orchestrator platform (see below).
A simple example might be deploying local recommendation models. These privately process a combination of data sources (e.g. age, location, interests, apps-on-device, browsing data and psychometrics) to provide Personal Intelligence which informs the serving of contextual coupons on local devices.
We have conceptualised methods to train models on-the-edge directly between devices utilising Privacy Enhancing Technology (PET) advances such as Secure Multi-Party Computation (SMPC). This is on our lab and development roadmaps.
2.3 Generative Intelligence
Running GPT LLM models on the total sum of all of a customer’s Life Data, holds enormous promise. There are a growing number of powerful open source ‘Small Language Models’ capable of being deployed on Smart Phones. These can be tuned on the local data held on the edge using techniques like Retrieval-Augmented Generation in the DataVault to create highly personal marketing.
However, for the moment at least, there are significant risks in deploying GPT LLM technology into a customer’s native device. A useful diagram from Gartner highlights where GenAI is and is not useful.
In addition to inherent biases in most foundational models, GPT LLM models are creative by nature, and therefore prone to hallucination. In some situations, such as content creation or movie suggestions, this presents a relatively low risk, but in others, this trait is highly dangerous and may leave the brand with major liabilities. For example, today, the technology may generate incorrect nutrition advice to a person with severe food allergies, or provide the wrong flight departure time to a travelling customer.
Consequently, the utility and readiness of Large and Small Language Models that run on-the-edge (on the device), with data also held on-the-edge, are being investigated and tested in the DataSapien Lab.
The Holistic Personal Intelligence Stack: Combining Three Tiers
We are excited by the results of our early experimentation that combines Generative AI with deterministic and probabilistic intelligence to provide optimal outcomes. We look forward to providing more details on this as our learning progresses.
This post is part two or three posts covering DataSapien Architecture. Part one us available here.
The long-form post covering all of the components of DataSapien Edge Architecture is also available here.
For more technical information on DataSapiens Edge Architecture, checkout the DataSapien Developer Portal.
If you have any comments or questions, please do get in touch. We’d love to hear from you.