APPLIED RESEARCH
Clockwork AI Lab
From small machines to autonomous construction intelligence
How much useful engineering intelligence can we build with the compute available to us?
The AI Lab is where Clockwork builds the engineering intelligence behind the consultancy. Each generation increased what the lab could do, not only how much it could compute.
The work is staged deliberately: start with the machine available, make the system modular and deterministic, and earn the next capability before claiming it.
Seven generations of lab
Each generation increased what the lab could do, not only how much it could compute. The capability line is the one that matters.
Generation 0
Clockwork
The first development machine, a conventional laptop.
Early AI experimentation under a hard compute constraint.
That constraint forced modular systems, structured information, deterministic methods and specialised agents.
Capability gainedAI-assisted engineering, local development, structured project data, first agent experiments.
Generation 1
Vesalius and Cirus
Vesalius was an early autonomous development and validation node.
Cirus was a Raspberry Pi event and communications spine.
Together they demonstrated distributed intelligence without one huge computer.
Both retired after hardware failures. The software architecture survived.
Capability gainedDistributed agents, event-driven systems, autonomous software development, machine-to-machine coordination.
Generation 2
Prometheus and Nexus
Prometheus runs reasoning, services and experimental cognitive systems.
Nexus was an early perception and domain-data node, turning documents and interactions into structured machine-readable knowledge.
Nexus retired after a hardware failure. Its functions were redistributed.
Capability gainedPersistent state, perception pipelines, domain knowledge, model-driven reasoning, autonomous workflows.
Generation 3
Local models and Vault
Clockwork was upgraded for larger local workloads and GPU acceleration.
Vault became the storage layer for engineering data, model traces, training examples, evaluations and artefacts.
Extensive open-weight model experimentation followed.
Capability gainedLocal inference, open-weight models, large-scale data collection, evaluation datasets, training-data flywheels, domain-specific model experimentation.
Generation 4
Clockwork2
Dedicated AI development compute, with dedicated GPU compute for local inference, autonomous development and parallel experimentation.
Engineering, project information, classification, programme, cost and evidence now operate against shared project state.
Capability gainedFaster local inference, autonomous development loops, larger agent systems, engineering evaluation, integrated project intelligence.
Generation 5
Leviathan
A dedicated Linux development and compute node for greater memory, concurrency and autonomous workloads.
The planned expansion to 256GB RAM is about more of the lab existing simultaneously: development agents, local models, evaluation harnesses, project simulations, engineering solvers, model routing, data processing and world-state systems.
The important unit is verified useful work per unit of compute.
Capability gainedHigh-concurrency autonomous development, larger local models, continuous evaluation, increasingly persistent systems.
Generation 6
PlannedTerminator2
The planned next compute platform for continuously operating autonomous systems.
Its purpose is to increase the rate of evaluating models, running experiments, specialising open models and simulating construction systems.
Capability gainedContinuous autonomous experimentation, model specialisation, large-scale simulation, physical-AI development, higher research throughput.
The important curve is capability, not hardware
Hardware is only useful when it turns into capability the lab keeps. Each generation feeds the next through one compounding cycle.
What the team watches is continuous capability improvement: the same loop run faster, on harder engineering problems, with the results retained.
- Build
- Test
- Observe
- Learn
- Retain
- Build again
From software into the physical world
The next capabilities leave the screen. Sensing, machine vision, position monitoring, material measurement and robotic control connect the engineering model to what is actually built.
Clockwork Engine provides the project and engineering intelligence. The sensing layer observes; the robotic systems act. Together they are the foundation of 3DP and of autonomous construction systems.
In development
Deterministic verification hardware
Given an explicit typed proposition, an implementation, a configuration and its inputs, the device proves whether the stated result follows. It returns one of four outcomes: verified, rejected, unresolved or error.
Interpretation stays with people and probabilistic agents. Deterministic verification moves progressively toward hardware, where a result can be checked without trusting the system that produced it.
The device is being developed in house. It is deliberately not named here while the work remains in development.
The lab's knowledge, in vector space
Structured knowledge and the source corpus, shown as interactive projections.
Fund or collaborate
The lab grows through funded research and engineering collaboration. If you fund applied construction research or want to build against these systems, get in touch.
Fund or collaborate