Activeloop developed a solution for processing and generating patents using enterprise-grade memory agents and their Deep Lake vector database. The system handles 600,000 annual patent filings and 80 million total patents, reducing the typical 2-4 week patent generation process through specialized AI agents for different tasks like claim search, abstract generation, and question answering. The solution combines vector search, lexical search, and their proprietary Deep Memory technology to improve information retrieval accuracy by 5-10% without changing the underlying vector search architecture.
This case study presents PatentGPT, an LLM-based solution developed by Activeloop in collaboration with Intel, AWS, and RCA AI. The presentation was given by David, founder of Activeloop, who previously worked on large-scale datasets during his PhD at Princeton University. The core problem addressed is the inefficiency of patent search and generation—with approximately 600,000 patents filed yearly and 80 million total patents globally, the traditional process takes 2-4 weeks to generate a patent and relies on outdated keyword-based search interfaces like the USPTO website.
The solution demonstrates how to build production-ready generative AI applications using what Activeloop calls “Enterprise Grade Memory Agents”—a multi-agent system that combines specialized LLMs, vector databases, and data infrastructure to handle complex patent-related tasks.
A key theme throughout the presentation is the gap between demo-quality applications and production-ready systems. David draws an analogy to self-driving cars: it’s easy to drive slowly in a neighborhood, but highway driving requires solving all edge cases—a process that took Tesla 7-8 years. Similarly, while it’s trivial to build a “shiny demo” on top of OpenAI APIs, the real competitive moat for companies lies in the data they collect and how they use it to specialize LLMs for their specific use cases.
The presentation argues that the current AI data stack is fragmented and inefficient. Companies typically have metadata in Postgres or Snowflake, images on S3, and now need a third tool (vector database) for embeddings. This creates a cumbersome workflow where data scientists must export data, link images, copy to machines, preprocess, train, run inference, generate embeddings, store in vector DB, and link everything back together. This iterative process is time-consuming and error-prone.
Activeloop’s Deep Lake is positioned as a unified storage layer that addresses these fragmentation issues. Key technical characteristics include:
The streaming capability is described as “Netflix for datasets”—data can be streamed directly from storage to GPU compute for training or fine-tuning, eliminating the need to copy and transfer data.
The PatentGPT system employs a meta-agent architecture designed for high fault tolerance and accuracy. When a user provides a query, the meta-agent decides which specialized agent should handle it:
Each agent is “well-scoped” with careful prompt engineering and access to the appropriate specialized model. This modular approach means each agent can be optimized independently, and the meta-agent serves as an orchestrator making routing decisions.
The workflow demonstrated includes:
A notable technical feature is the automatic generation of filters from natural language queries. When a user asks for “patents from 2007,” the system recognizes that “2007” should not be part of the embedding search (since dates carry no semantic meaning in embedding space) but should instead become a filter condition. This is automated rather than requiring manual specification or explicit agent configuration—the query engine parses the intent and applies appropriate filters before running vector similarity search on the filtered subset.
The demonstration shows several modes of interaction:
Perhaps the most technically significant claim in the presentation is around “Deep Memory,” a feature designed to improve retrieval accuracy without changing the vector search operation itself. The presenter shares evaluation metrics on a dataset, comparing:
The key insight is that better indexing—learned from query patterns—can improve question answering accuracy. For RAG applications, the top-K recall (whether the correct document appears in the top 10 results) is critical because if the right context isn’t retrieved, the LLM cannot produce accurate answers.
The presenter emphasizes that many RAG solutions work “70% of the time” and the challenge is pushing accuracy above 80-90%. Deep Memory is positioned as a way to improve accuracy “out of the box” while still allowing additional techniques like hybrid search and re-ranking to be layered on top.
The presentation touches on several production-grade concerns:
The architecture is described as analogous to computer memory hierarchy: smaller context LLMs as L1/L2 cache, larger context LLMs as L3, the memory API layer (LangChain, LlamaIndex, agents operating on vector databases), and Deep Lake serving as both the underlying storage and the training data source for fine-tuning.
While the presentation makes compelling claims about Deep Memory’s accuracy improvements and the benefits of unified data infrastructure, it’s worth noting that:
The Tesla/self-driving analogy, while illustrative, somewhat oversimplifies the challenges—patent generation has different risk profiles than autonomous vehicles, and the 7-8 year timeline isn’t directly applicable.
Activeloop offers a free certification course at learn.activeloop.ai in collaboration with Towards AI and Intel, covering practical projects for building generative AI applications with LangChain and vector databases.
Snorkel developed a specialized benchmark dataset for evaluating AI agents in insurance underwriting, leveraging their expert network of Chartered Property and Casualty Underwriters (CPCUs). The benchmark simulates an AI copilot that assists junior underwriters by reasoning over proprietary knowledge, using multiple tools including databases and underwriting guidelines, and engaging in multi-turn conversations. The evaluation revealed significant performance variations across frontier models (single digits to ~80% accuracy), with notable error modes including tool use failures (36% of conversations) and hallucinations from pretrained domain knowledge, particularly from OpenAI models which hallucinated non-existent insurance products 15-45% of the time.
Notion AI, serving over 100 million users with multiple AI features including meeting notes, enterprise search, and deep research tools, demonstrates how rigorous evaluation and observability practices are essential for scaling AI product development. The company uses Brain Trust as their evaluation platform to manage the complexity of supporting multilingual workspaces, rapid model switching, and maintaining product polish while building at the speed of AI industry innovation. Their approach emphasizes that 90% of AI development time should be spent on evaluation and observability rather than prompting, with specialized data specialists creating targeted datasets and custom LLM-as-a-judge scoring functions to ensure consistent quality across their diverse AI product suite.
DoorDash faced challenges in scaling personalization and maintaining product catalogs as they expanded beyond restaurants into new verticals like grocery, retail, and convenience stores, dealing with millions of SKUs and cold-start scenarios for new customers and products. They implemented a layered approach combining traditional machine learning with fine-tuned LLMs, RAG systems, and LLM agents to automate product knowledge graph construction, enable contextual personalization, and provide recommendations even without historical user interaction data. The solution resulted in faster, more cost-effective catalog processing, improved personalization for cold-start scenarios, and the foundation for future agentic shopping experiences that can adapt to real-time contexts like emergency situations.
