The Key 4 AIThe craft

Decide. Build. Prove. Run.

What happens when someone brings us a problem, in the order it happens. The fourth movement is where most teams stop being useful.

Decide Agents

Craft · Decide

Most problems are not language-model problems.

The first skill is knowing what kind of intelligence a job actually needs. Vision, language or structured data. A classifier or a generator. A model at all, or a rule that somebody should just write down. Most failed AI projects did not fail at the model. They chose the wrong family for the problem, usually a large language model where a small classifier would have been righter, more cost-effective and testable.

A problem arrives, and splits three ways.

Vision

Inspection and defect detection
Document and image understanding
Anomaly detection in visual streams

Language

Extraction and classification
Retrieval over private corpora
Generation only where output is checkable

Structured

Scoring, ranking and forecasting
Anomaly detection over records
Tabular and time-series modelling

Choosing correctly is the capability. Building it afterward is the easier part.

Craft · Build · Deep learning

Real deep learning, for the specific problem in front of you.

Most of the intelligence an enterprise or a mid-sized company needs is not a language model. Inspection on a production line. Anomaly detection over a million transactions. Forecasting over years of records. Document understanding at volume. These are deep learning problems, and we build them outward from the problem, not inward from a general model.

What we build

Convolutional, recurrent, transformer and graph models designed for the task. Not a general model trimmed to fit, which is still shaped by what it was originally for.

Where it runs

On the hardware already there, at the edge, on the plant floor, with no network required and nothing leaving the site.

What makes it ours

Architecture research: designing the model around the job. That is where the smallness and the speed come from, and it is research, not configuration.

This is the larger branch of the build, not the smaller one. Most of what an enterprise needs lives here.

Craft · Build · Small language models

We make a small model expert at one job.

Where language is the right tool, we make it expert rather than general. We have run this end to end and taken a base model to a domain expert with it.

A base model

Generally capable. Knows a great deal. Reliable at nothing in particular.

A verifier

A machine that can tell a right answer from a wrong one in this domain. The hard part, and it has to be built first.

A specialist

Reinforcement learning against that verifier until the model is excellent at one job and small enough to run anywhere.

The verifier had to exist before the model did. Hold that thought. It is the next section, and it is the whole company.

Craft · Prove

The machine that decides what is correct.

Everything on the preceding pages depends on this one. To train against a verifiable reward you need a verifier. To prove a model is safe to deploy you need the same machine. And once it exists, it does not care whose model it is grading.

Our own modelwe built it

It becomes training.

The model gets better at the one job, measurably, and we can show the measurement.

The verifier

Anyone else's modelwe had no hand in it

It becomes assurance.

The same measurement, on a model we had no hand in building and no interest in flattering.

One role per model. We build it, or we assure it. Never both. You choose which before work starts.

Building and proving are one piece of engineering here, so independent assurance does not cost us what it costs everyone else.

The Bench

Craft · Run · Agentic systems

Put it to work. Keep it working. Agents included.

Agents

Multi-agent workflows and swarms
Orchestration and tool use between agents
Harnesses for problems with no single right answer

Deployment

On premise and at the edge
Compressed to fit the device you have
Nothing crossing a boundary to be processed

Operation

Monitoring and drift detection
Incident response on short clocks
The unglamorous years after launch

Deployed and then operated, rather than handed over and forgotten.

Craft · Run · Agents

An agent you can hold.

An agent is a model that acts. It calls tools, moves through systems, and changes things. That makes the question of proof sharper, not softer: an answer that is wrong can be ignored, an action that is wrong has already happened.

Orchestration

Agents decomposed by role, coordinated by a plan that is inspectable, not by a prompt that is hoped for.

Tools and boundaries

Every tool an agent can call is enumerated, permissioned and logged. Nothing crosses a boundary it was not given.

Harnesses

Agent runs are replayed against harnesses that grade the outcome, the path and the cost, so an agent can be improved on evidence rather than on anecdote.

The verifier, again

The same machine that grades a model's answer grades an agent's action. One role per agent: we build it, or we assure it. Never both.

Agents that act inside your walls, on your hardware, under a verifier that grades what they did.

Honest enough to test.

A value. We claim nothing you cannot check for yourself, and we never grade our own work.

One address

One sentence about what you need to hold.

info@thekey4.ai