DeepStructure's Mission: Unlocking the Next Wave of AI Applications

We started DeepStructure because we were frustrated by the difficulty of building production-ready AI applications.

What does it mean to have a production-ready AI system? To us, it comprises:

• Basic AI infrastructure for LLM inference, tool calls, retrieving relevant context

• Connections to your data sources such as Google Drive and databases, in order to supply fresh data and documents to LLMs — and keep the data fresh as it changes over time

• Connections to your human interfaces such as Slack and email

• A feedback loop to know how your system is performing in the field

To many developers, these are new, daunting challenges — AI applications don't look like the typical web applications we cut our teeth on.

Building with LLMs and semantic search is more like working with flaky humans than deterministic request-response services. Sometimes they do what you want, sometimes they fail or generate garbage. Sometimes they're degenerately slow, and sometimes you hit rate limits. It's a challenge just to keep track of the sequence of operational steps and data flows in a given workload.

And perhaps most daunting of all, AI technology has been evolving at a rate we've never seen before. It seems like every week there's a new breakthrough in foundation models, iterated algorithms, prompting techniques, … it can be too much!

DeepStructure's mission is to solve the hardest of these problems for you, so that you can focus on building the features that your users want. Here's what we provide for you:

AI infrastructure as an API

You should be building at the level of Assistants, Files, Agents, and Workflows — not worrying about the details of embeddings, vector dimensions, task orchestration, etc. You can come down to the engine room and tweak these when you need to, if you ever need to.

Data sources as declarative components

As a developer, you should be able to declare how your system should react to events from the outside world — files changing, new messages arriving, timers firing, etc. And you should be able to observe the chain of events that produces a new response or output. But do you really want to build all that plumbing yourself — and then battle-harden it against intermittent flakiness?

Your accountant has this experience working with spreadsheets. Your marketing department has it building no-code automations. Why can't developers have this for "real code" too?

Compatibility with your existing systems

You can start benefitting from DeepStructure by changing a single line of code in your current client SDK. Your existing services can communicate with DeepStructure via standard, secure protocols like HTTPS and WebSockets, so that you can for example seamlessly interoperate with your Next.js app. And DeepStructure has prebuilt components for common human-machine interfaces to connect with your users wherever they are.

Built-in feedback and support for evals

It's hard enough to write high-quality evals for the tasks your application is trying to do. You should focus on the evals, and not the machinery of how your datasets are labeled and curated, or how experiments are run.

Escape hatch for power users

At its core, DeepStructure is a TypeScript engine for reliably processing data and events. The high-level conveniences we provide to developers, such as Assistants, build on lower-level primitives of blob stores, reactive events, tables, and arbitrary TypeScript workflows that developers are encouraged to use directly when it suits them.

And one more thing …

What DeepStructure is actually helping you do, but not making a lot of noise about, is to build a data flywheel for your AI system. With a data flywheel, the more data that flows through your system, the better it becomes — and thus more data will flow through it with increased usage.

And when you have a data flywheel in place — with the right evals — we want to put your optimizations on autopilot. Then you can stop fearing change and learn to embrace it: you'll be able to deploy new tech with confidence before your competitors do.

—————

We've intentionally kept this mission statement at a pretty high level. We'll be sharing more posts about the specific features we're building for you, and how to use them. And expect things to evolve quickly …