Saddle The Bull

You've likely spoken to an AI by now. Tools like ChatGPT are everywhere, generating startlingly human text. These large language models (LLMs) are masters at predicting the next word, trained on oceans of data. They're incredible pattern-matchers, perfect for an open-ended chat or reflective discussion.

Now, try getting one to follow strict rules. Ask it to generate data that must fit a precise format, interact reliably with another piece of software, or execute a specific, structured action. The brilliant conversationalist stumbles, putting information in the wrong place, hallucinating details, mangling formats.

This is the wall you hit when trying to build dependable systems with what is essentially an improvisation machine. To build reliable AI, you need more than just powerful models. Getting the structure right, every time, is a prerequisite for building anything meaningful on top.

Today, I'm introducing the Proxy Structuring Engine (PSE).

It’s the engineered solution to that friction, that wall - the fastest, most reliable way to generate structured outputs from modern LLMs. The PSE did not come from market research: it was born in the debugger, tracing why systems built on LLMs kept breaking.

It is the solution I needed as a builder to push beyond simple chatbot interactions. The goal was always systems that could execute, and that requires structure and output you can always depend on.

The challenge is that LLMs are probabilistic; meaning they suggest many possible next words. Older libraries try to approximate structure by masking a few obvious "bad" words or by validating structure after the fact. That is slow, costly, and fails to scale to complex or nested formats where validity depends on context deep within the structure. The brittle nature of these approaches is why I built the PSE.

This meant recognizing that structure implies hierarchy. A simple linear check isn't enough.

My own approach for complex, layered problems is: find a path forward; if there’s an issue, fix that first, then continue forward. This internal model was the inspiration for the PSE: it natively understands nested structures and recursion - navigating between different sub-states. This layered state tracking, navigating up and down the structural hierarchy as needed, is what sets the PSE apart from other, flatter, approaches.

The Proxy Structuring Engine is a new way to think about and build with LLMs. It allows you to generate reliably structured output, even when the underlying LLM is probabilistic. The PSE uses a lightweight layered state machine architecture to track the precise state of the structured generation. It simultaneously tracks multiple valid states, dynamically filtering invalid choices and actively guiding token selection to ensure continuous structural validity. This stateful approach is what allows the PSE to handle arbitrary complexity consistently, where other methods simply cannot compete.

Critically, the engine is decoupled from the specific LLM doing the generation. The PSE processes and refines the model's raw output via standard interfaces, working directly with the model to choose the next token. It does not affect the model's predictive or creative capabilities; simply guiding the internal generation logic.

This model agnosticism is deliberate. My goal was an infrastructure layer that provided a consistent guarantee - whether the underlying model is open source, proprietary, large, or small.

The PSE is the first piece of technology from The Proxy Company, built out of necessity for the AI systems I believe are the future. It's engineered for robustness and designed to give builders the control they need. Reliability no longer depends on hoping your model behaves; it's guaranteed by the engine itself, allowing builders to choose the best LLM for their needs without sacrificing structural predictability.

To see the full breakdown on what the PSE enables – predictable execution, dependable outputs, truly robust AI systems – head over to the main Proxy Structuring Engine page.

The core Python library is open source and production-ready. Read the official technical documentation or leverage our paid services and support.

Getting the structure right is only the first step.