Predictika granted second patent
Predictika has been granted a very broad-based second patent (see first patent) for conversational AI Agents.
US Patent # 11,914,970 B2, titled “System and Method for providing a model-based intelligent conversational agent” was granted to Sanjay Mittal (http://www.linkedin.com/in/sanjaymittal) and Awhan Patnaik, who developed the technology and software while at Predictika, a Fremont, CA based startup doing pioneering work in conversational AI.
The core of Predictika’s two patents is a deterministic Logic Validation Engine that we believe is a key component in the development of LLM-based AI Agent applications. It is widely accepted that LLM chat tools are not very good at reliably following business logic in carrying out their tasks. Predictika’s Logic Validation Engine is one good way to not only continually check the results returned by the LLMs to ensure that it follows the task logic, but also proactively prompt the LLM for the next step in a multi-step application.
Key features of Predictika’s Logic Engine: -
- A task model driven approach where a declarative task model is used to embed sophisticated domain knowledge and business logic for carrying out the task. The task model is object-oriented with constraints and rules that define relationships between model parameters.
- The models can be built using our declarative modeling language or created from structured data. This has been used for many of our applications such as for hospitality (see demo version of the applications being deployed), education, and field service AI agents.
- The models can also be created using an intermediate translator when the external data needs more processing, usually in multiple passes. This has been done for processing restaurant menus to automatically create models for food ordering AI agents (see half and half pizza order demo,
multiple items order demo, or
general questions, order, cross-sell demo).
- Rules tend to be procedural while constraints are more declarative. By allowing both we can model in the most appropriate way.
- A wide variety of constraint types (e.g., algebraic expressions, set and list expressions, tables, graph relations, temporal relations) are built-in and new ones can be defined rather easily.
- A parameter (which has some overlap with slots in NLP) in addition to actual value(s) can also maintain possible values that result from propagating constraints from known to unknown parameters.
- A tree-structured graph of choice states is maintained where a state in this graph corresponds to a ‘free choice’ by the user.
- The Logic Engine can be invoked via a Python API or via a natural language API since the underlying model parameters also map to NLP entities.
- The Logic Engine is responsible for managing the state of the model parameters as it is given incremental information about user choices. It can revise the state or maintain alternative possible states.
- Based on the known parameters, the Engine continually makes inferences using the relevant constraints and rules.
- The Logic Engine also manages the conversation flow with the user based on the dynamic nature of the interaction with the users. This includes managing the evolving context for multi-step dialogs.
Predictika is also making it's patents and software available for licensing so everyone in the industry can benefit from our path-breaking technology. Please contact Laurie Spoon at laurie@predictika.com or 408-438-7261