US 12142371

US Patent 12142371, titled "Low-latency conversational artificial intelligence (AI) architecture with a parallelized in-depth analysis feedback loop," was issued to Healthgpt Inc. dba Hippocratic AI.

Summary of US Patent 12142371

• Title: Low-latency conversational artificial intelligence (AI) architecture with a parallelized in-depth analysis feedback loop
• Assignee: Healthgpt Inc dba Hippocratic Ai
• Inventors: Munjal Shah, Vishal Parikh, Meenesh Bhimani, Subhabrata Mukherjee, Alex Miller, Saad Godil, Debajyoti Datta, Paul Gamble, Rae Lasko
• Filing Date: February 29, 2024
• Issue Date: November 12, 2024
• Abstract: The patent describes a low-latency conversational artificial intelligence (AI) architecture that includes a primary AI, such as a large language model (LLM), engaged in a conversation with a human. Concurrently, a "second opinion module" performs a more in-depth analysis of information provided by the human. If this second opinion module determines that a response from the primary conversational AI needs clarification or expansion, it provides feedback that the primary AI uses to offer a clarification to the human during the conversation. This parallel processing reduces latency, making the conversation more natural. Additionally, a data extraction module operates in parallel to extract facts from human responses, creating a searchable conversation summary, such as a knowledge graph, which the conversational AI can quickly access to reference previously provided information, further reducing latency.

Plain-Language Overview of Independent Claims:

The patent contains three independent claims (Claims 1, 5, and 17).

• Claim 1 (Multi-turn conversational system with control logic): This claim describes a multi-turn conversational system comprising a first large language model (LLM) that acts as a conversation interface to conduct human-like conversations with a user over multiple turns. This LLM has been extensively trained (over one thousand gradient update iterations). The system also includes a "control logic" that communicates with the LLM. This control logic evaluates the ongoing conversation and generates "control signals." These signals help shape the future responses of the LLM, influencing the subsequent turns of the conversation with the user. The control logic may include specific sub-components such as trigger detection, question insertion, and answer classification logic.
• Claim 5 (Retro-improving conversational system with retro-improvement logic): This claim details a system that improves its past responses. It features a large language model (LLM) based conversation interface, also trained with over a thousand gradient update iterations, that engages in multi-turn human-like conversations. During these conversations, human-machine response pairs are generated. The system includes a "retro-improvement logic" that monitors these past exchanges. If this logic detects that a previous machine response needs to be improved (e.g., clarified, corrected, or expanded), it generates an "improved machine response" and causes it to be presented to the user. This improvement can happen at a later point in the conversation, potentially after other exchanges have occurred. The retro-improvement logic can analyze the past exchange, select specialized "specialist engines" (which can be smaller LLMs) relevant to the conversation, and these specialist engines can generate the improved response or execute follow-on tasks.
• Claim 17 (Retro-improving conversational system with conversation interface and retro-improvement logic): Similar to Claim 5, this claim also describes a retro-improving conversational system with a conversation interface that engages in multi-turn human-like conversations, including human-machine response pairs. It also features a "retro-improvement logic" that detects when a previously expressed machine response requires improvement. The retro-improvement logic generates an improved machine response and causes it to be expressed to the user. This claim is broader than Claim 5 in that it does not explicitly specify the conversation interface as an LLM trained with a certain number of gradient update iterations, allowing for other AI implementations.

CAFC 2026 Dockets:

A search of CAFC 2026 dockets for patent number 12142371 did not return any direct results. However, it is noted that a Post-Grant Review (PGR) case, PGR2025-00075, was filed against US Patent No. 12,142,371 B1 by OSSI Corporation with the Patent Trial and Appeal Board (PTAB), challenging claims 1 through 20 as unpatentable. This is a PTAB proceeding, not a CAFC docket entry. The legal status of the patent is listed as "Active".