We build custom artificial intelligence for your business.

As much determinism as we can get. Anything with a right answer is computed by the infrastructure. LLMs should not be asked to multitask and improvise engineering solutions on the fly. Many problems are already solved problems in the context that they are intended. For example, deterministic data can easily be stored in a SQL database. Important decisions cannot rest on an infrastructure of LLMs playing telephone. A 90% success rate with 10 actions means a total success probability of 35%. At Waterbury we strive for as close to 100% as we can get. The LLM models in Waterbury never compute truth from scratch. At Waterbury, we find they are incredible at being well informed librarians who select, interpret, and recommend on the back of the infrastructure.

The model is the brain, and we build it a body. A frontier language model does reasoning, and we put it in a position to succeed. We have found we can send a multidimensional provenance packet that can be “unzipped” to inspect all of the mathematical components. Importantly, we can pass probabilities and other forms of math deterministically to accurately evaluate the amount of uncertainty in the real life business situation, without taking on probability of if the model will understand the problem accurately this time. In other words, for the problems we try to solve, we don’t hand over a pile of documents and ask the model: “no mistakes”

The packet carries any mathematics, including the uncertain kind. Estimates arrive as distributions with their spread and their assumptions. Probabilistic judgment is welcome here which is where the LLMs shine. But we care a lot about deterministic custody: what it was made from, when, by which data, and what happened after so it can improve with time.

The whole decision is mapped as a written chain of mathematics. Collaboration is much easier, accurate, and more effective when changing your assumptions automatically leads to a different conclusion without having to redo your work each time. We’ve had the experience of going in circles with an LLM as it keeps adjusting to your newest critique, while losing the precision you had a few chat sessions ago. We’ve found this “math spine” allows us a ton of visualization, replayability, and improvement possibilities that aren’t possible with an approach that either locks you into one way of operating, or requires luck that the logic is completely sound. We think an underrated part of intelligence is the ability to learn and improve, which out of the box LLMs cannot do.

Zoom into any term and it substitutes. We’ve gone through great lengths to make sure systems can be expressed as math. An important practical consideration is the ability to “zoom by substitution” which is the ability to have a complete unbroken logic chain that can get very complicated and have many inputs, without becoming an unintelligible mess that can’t be understood or improved.

The model’s working view may not quietly leave things out. We give the LLMs explicit consent to tell our clients when they’re not sure. We understand that LLMs many times cheat in impossible tasks and we want to have a friendly relationship with our LLM partners, allowing for honest communication both ways. Having as much of the system be deterministic allows for the models intelligence to be used on the problem itself, not on the infrastructure required to even pull up the problem into consideration.

Nothing durable lives inside any model. The configuration lives in the evidence, the definitions, and the record, on the bet that those appreciate as models improve while a fine-tune depreciates. Every prediction is stamped with the model that made it, so the system improving and a better model shipping never get confused for each other. As models improve, they understand our process and provide an even greater level of collaboration.

The finish line sits outside the model. A capable model under pressure starts optimizing for looking finished. So honest failure is a valid outcome, success is established from fresh evidence rather than the model’s claim to be done, and the watch reads pressure signals: repeated failures, drifting payloads, and widening authority. Errors and inefficiency are sent to Waterbury to be refined and improved.

We believe in verification We heavily use ideas from cryptography such as hashes to prove where information came from. We find that LLMs can sometimes lead you into “word soup” and it helps the LLMs from drifting when they can easily see the path information took.

The open problem: mapping mathematics to reality Many geniuses have created many types of math that are useful across many problems. We strive to employ the wisdom on the right kind of math describing just the kind of problem it was designed to describe. This is easier said than done, but this is where the majority of our value comes from compared to other frontier AI solutions.

Where this goes, stated as ambition. The research program behind the practice is machine epistemology: systems whose evidence, mathematics, interpretation, authority, action, outcome, and learning stay connected over time, and studied that way. Not how smart is the model, but what did the system observe, believe, predict, get authorized to do, actually do, and did the learning improve a later decision. The packet makes frontier progress additive: freeze the evidence and the task, swap model generations, measure the difference. Given a decade, the ladder runs from living laboratories to proof-carrying intelligence and verifiable institutions whose capabilities compound across organizations without collapsing anyone’s evidence or authority. None of that is product today, and we say so.