Baseplate - Heroku for LLM Apps
Baseplate is a platform that simplifies backend development for LLM apps by offering a unified database, user-friendly APIs and a spreadsheet-style interface for data management.
Note: Attack Capital is an investor in Baseplate through our Demo Day Fund.
TLDR;
Baseplate wants to be the "Heroku for LLMs," providing simple APIs for developers to connect LLMs to an ever-changing set of data, manage multiple databases, and optimize the backend for their LLM apps the way Heroku did for the cloud.
The core of Baseplate is its simplified multimodal database, which allows users to store text, embeddings, data, and metadata in one place. Devs can edit vectors and metadata through a spreadsheet-style interface.
Baseplate also has an App Builder where users can iterate on prompts with input variables, create context variables that pull directly from a dataset at query time, and configure exactly how the search is performed and how it is integrated with the prompt.
The founders, Ani and Andrew have known each other since middle school. Baseplate was started as a side project while Ani was working at Logitech as an Associate Product Manager and Andrew at Google X as a Software Engineer.
From middle school classmates to now Co-Founders, Ani Gottiparthy and Andrew Luo are building Baseplate, a platform that allows LLM (Large Language Model) app developers to store, manage, and consume multimodal data from a single set of APIs.Â
When Ani and Andrew joined Y Combinator’s Winter 23 Batch, they were working on a no-code solution that would fine-tune, monitor and manage LLMs for users who could use LLMs apps effectively to save cost and time.
Soon they realized that most of the AI model providers, like Open AI, would fine-tune and optimize LLMs for their users with time as the technology advances - making Baseplate obsolete. Undeterred, after some discussion with their batchmates and the YC team, they locked down a much more fundamental problem to solve - building a unified backend for LLMs.
THE NEED FOR BASEPLATEÂ
Baseplate wants to build Heroku for LLMs. Heroku was started to give developers the freedom to focus on their core product without the distraction of maintaining servers, hardware, or infrastructure. Baseplate wants to do the same for developers who are building LLM apps, which means providing simple APIs for - data source integrations, async embedding jobs, vector databases, bucket storage for nontextual data, and potentially an additional database for the text data.
In most applications, LLMs need to be connected to a constantly changing data set. If the data is simple as a couple of PDFs and the model just needs to parse them, then it is manageable. However, when developers are working with large, multimodal datasets that consistently need to be updated, re-indexed, and sometimes replaced, then they are in for a tedious and painstaking job. Developers also have to manage multiple databases - one for vectors and one for their other data - another tedious process on top of a tedious process.
Baseplate’s simple no-code spreadsheet-like interface lets developers optimize the backend for their LLM Apps This also means they don’t have to maintain and manage separate databases for vectors and regular data anymore. Engineering teams can simply use their multimodal context database to build rich user experiences with LLMs.
LOGISTICS OF BASEPLATE
The core of Baseplate is its simplified multimodal database, which allows users to store text, embeddings, data, and metadata in one place. Baseplate's spreadsheet-style interface lets developers edit their vectors and metadata (a super complex process with existing tools) and add images that can be returned at query time. They can also choose between standard semantic search or hybrid search (weighted keywords/semantics for larger, more technical datasets).Â
Hybrid search on Baseplate utilizes two open-source models that can be tuned for use cases (instructor & SPLADE). Datasets are organized into documents, which users can keep in sync through the API or through the UI (this way, users can keep their datasets fresh when ingesting data from Google Drive/Notion/ etc).
After the datasets are set up, Baseplate has an App Builder where users can iterate on prompts with input variables and create context variables that pull directly from a dataset at query time. The platform gives all the knobs and dials so that users can configure exactly how the search is performed and how it is integrated with the prompt.
When users are satisfied with an app configuration, they can deploy it to an endpoint. All users need is a single API call, and the platform pulls from one (or multiple) datasets in the user’s app and injects the text into the prompt.Â
The platform also returns all the search results in the API response so that users can build a custom UX around images or links in their dataset. Endpoints have built-in utilities for human feedback and logging. With GPT-4 being able to take images as input, Baseplate will soon be working on a way to pipe images from your dataset directly to the model. And all of these tools are in a team workspace, where users can quickly iterate and build together.
The platform also returns all the search results in the API response so that users can build a custom UX around images or links in your dataset. Endpoints have built-in utilities for human feedback and logging. With GPT-4 being able to take images as input, Baseplate will soon be working on a way to pipe images from your dataset directly to the model. And all of these tools are in a team workspace, where users can quickly iterate and build together.
THE BASEPLATE TEAM
When Ani and Andrew were together in Middle School, none of them would have thought that they would start a company together some ten-odd years later.Â
After high school, they both went their separate ways, Andrew to UC Santa Barbara and Ani to UC Berkeley, and then Andrew joined Google X as SWE and started working on data infra and integrating knowledge graphs with LLMs. While Ani joined Logitech as an Associate Product Manager and started working on their Computer Vision and Machine Learning offerings. They worked on Baseplate as a side project first and left their job the moment they got their YC acceptance in December 2022.
MARKET OPPORTUNITY AND THE ROAD AHEAD
Companies across sectors are recognizing the potential of large language models (LLMs) to enhance their offerings. Both Software-as-a-service (SaaS) and consumer companies can integrate large language models (LLMs) for intelligent automation, personalized interactions, and content recommendations.
However, integrating LLMs presents challenges in data management and model integration. Baseplate addresses these challenges with a unified backend and intuitive tools, simplifying LLM integration for developers.
As demand for LLM-enhanced solutions grows, Baseplate is positioned to facilitate AI-powered transformation for developers and organizations. Led by Ani Gottiparthy and Andrew Luo, Baseplate aims to unlock the full potential of LLMs through a simple no-code solution and make it simple for developers to build LLM apps without the distractions of constantly managing the backend.
