Clockwork Consultancy

Construction intelligence · applied research

From rented intelligence to owned construction capability.

Clockwork is building an in-house computational intelligence capability for construction.

The AI Lab is where it is developed, tested and extended: autonomous software development, local and open-weight models, model training and fine-tuning, persistent project state, deterministic verification and our own verification hardware.

The Lab is the shared capability layer enabling our consulting and driving 3DP robotics and sensing.

We use frontier intelligence where it earns its place.

We own the construction knowledge, data, evaluations, orchestration, engineering methods, verification systems and physical interfaces that compound around it.

Every project should leave the system more capable than it found it.

Autonomous developmentLocal AIModel trainingEngineering world stateDeterministic verificationDeterministic hardware

Seven generations. One direction.

What started as a browser session became an engineering system.

Seven generations. Each retains the capability of the last. Nothing resets; capability accumulates.

The first generation used web ChatGPT. Then came multiple models, coding tools and small local models. Machines became connected. State began to persist. Evidence stopped disappearing between sessions. Local models became useful. Autonomous development became practical. Real construction projects began producing reusable data, evaluations and engineering knowledge.

Now the system operates across local and frontier models, autonomous coding, persistent project state, governed evaluation and verification. The next step takes that capability into specialised models, dedicated hardware and the physical world.

ChatCopilotsToolsAgent teamsOpen weightsGoverned specialisationPhysical AI

We are not trying to eliminate frontier AI. We are making sure the valuable part of the system belongs to Clockwork.

The forensic record

Generation 0

Hardware
ClockworkFirst constrained development machine.conventional laptop
Models
Web ChatGPT, GPT-3.5 era.
Tools & harness
None
Data & evidence
None
Capability
Structured
Milestone
Structured methods forced by constraint.

Generation 1

Hardware
Vesalius + CirusAutonomous-development node and event spine.Vesalius 2-core, 8 GB · Cirus Raspberry Pi 4, 8 GB
Models
+ Gemini, ChatGPT-4; local Gemma 3 1B under a hard compute limit.
Tools & harness
Cursor · Gemini CLI, early agentic coding.
Data & evidence
Distributed event log.
Capability
Distributed
Milestone
Event spine makes the machines one system.

Generation 2

Hardware
Prometheus + NexusPersistent reasoning and domain/perception nodes.Prometheus 16 GB, integrated graphics · Nexus 16 GB
Models
Frontier APIs continue alongside local experiments.
Tools & harness
Early autonomous development systems.
Data & evidence
Persistent state, perception, beliefs.
Capability
Persistent
Milestone
Reasoning and state persist between runs.

Generation 3

Hardware
Clockwork + VaultUpgraded local compute and the storage layer.i7-10750H, 24 GB, GTX 1650 Ti · Vault CPU-only
Models
Claude and Claude Code; local models to ~4B.
Tools & harness
Claude Code.
Data & evidence
Vault, raw capture → curated evaluation and training material.
Capability
Retained
Milestone
Evidence is retained rather than discarded.

Generation 4

Hardware
Clockwork2Dedicated local AI compute.Ryzen 7, 64 GB DDR5, RTX 4050 6 GB
Models
Qwen 3 8B; Qwen 2.5 14B-class tried, compute-limited; Kimi Code.
Tools & harness
Kimi Code.
Data & evidence
Training-data flywheel; shared project state.
Capability
Integrated
Milestone
Engineering, cost, programme and evidence on shared state.

Generation 5

Hardware
LeviathanConcurrency and larger local workloads.Xeon W-2145 8C/16T, 32 GB (256 GB ceiling), RTX 3060 12 GB
Models
DeepSeek (cloud) + Ollama day-to-day; fully off Anthropic; selective Codex.
Tools & harness
opencode · Clockwork Coding Harness (experimental).
Data & evidence
Provenance, traces, verification ratchet.
Capability
Autonomous + Governed
Milestone
Outputs must earn authority; deterministic verification begins.

Generation 6

Hardware
Terminator2Planned, demand-led.no locked specification
Models
Specialised capability teams; Codex retained for the hardest planning.
Tools & harness
Clockwork Coding Harness, internal design-stage.
Data & evidence
Physical feedback loop.
Capability
Specialised + Physical
Milestone
Continuous operation; physical reality returns evidence.

The compounding advantage

Work becomes evidence. Evidence becomes capability.

Every useful piece of work can produce structured project state, traces, decisions, measurements, corrections and outcomes. That evidence is retained, curated and evaluated. Selected capability can then be trained, specialised, tested and promoted back into live operation.

Access to powerful models is not the moat. The compounding system around them is.

Every cycle leaves behind more data, stronger evaluations and more reusable capability.

The lab builds the products. The products build the lab.

Clockwork Engine produces engineering evidence. Autonomous development improves the software that produces it. 3D printing, sensing and robotics will return evidence from the physical world. Three loops. One compounding system.

* Engineering deliverables pass two human gateways before issue. AI may propose, calculate and produce; humans retain approval authority.

Operating today

This is already considerably beyond prompting a chatbot.

Autonomous software developmentUsed internally: coding agents can plan, execute, test, review and continue through governed engineering work.
Training and fine-tuningWe train and fine-tune models on Clockwork data and task-specific datasets, using local compute where practical and external GPU compute when scale demands it. Candidate models are evaluated against Clockwork tasks before promotion.
Local and open-weight inferenceSelected workloads run on Clockwork compute.
Model evaluationModels are tested against Clockwork tasks rather than selected by brand.
Persistent project stateOperational within Clockwork systems: engineering, cost, programme and source material share one state.
Selective frontier routingHigher-capability external models are routed where they provide sufficient value.
Retained evidenceOperational work produces retained data and evidence, not disposable conversations.
Verification before authorityOutputs can be tested against deterministic methods before they are allowed to carry authority.

The objective is not to own every model. It is to own the compounding system around them: what they learn from, how they are evaluated, what they are allowed to do, what becomes reusable capability, and what is trusted enough to carry engineering authority.

Not another chatbot

LLMs reason. Clockwork remembers, computes and verifies.

Large language models remain extremely useful. They interpret, search, reason, propose and communicate. But a construction project cannot live inside a chat window.

Clockwork Engine maintains a persistent computational representation of the project: geometry, engineering logic, programme, cost, evidence and decisions. Agents can propose changes. The World State Engine can represent those changes. Engineering systems calculate their consequences. Deterministic systems can verify selected claims. Humans retain authority.

World State Engine

Structured · persistent · grounded · for real projects

Project state
Geometry
Cost
Evidence
World State Enginepersistent · deterministic
Programme
Engineering logic
Persistent stateDeterministic structureEvidence-linkedProject intelligence
State → change → consequence → evidence → decision → updated statePersistent engineering intelligence.

LLM

Flexible · general · probabilistic · conversational

How can I help?
Summarise this…
Draft a response…
Consider the trade-offs…
LLMprobabilistic · general
That depends…
In general…
Prompt-basedProbabilisticContext-limitedGenerative reasoning
Prompt → answerUseful general reasoning.

A language model generates a response. Clockwork maintains, changes and reasons over a computational model of the project.

Beyond software · in development

We are developing a deterministic verification substrate in hardware.

Given admitted inputs and a defined proposition, it independently checks whether the claimed result follows. It does not trust the model or the software that produced the claim.

This creates a path from probabilistic intelligence to authority-bearing engineering computation without giving a learned model authority over correctness. The same principle can ultimately operate inside the World State Engine, optimisation systems and physical machines.

Software verificationAuthority-bearing receiptsVerified optimisationWorld-state integrationPhysical systems

The direction is deliberate: verification moves from a software promise to a property of the computational stack.

One loop, three systems

Three operating loops

The compounding loop runs three times: in software, in engineering and in the physical world.

Three operating loops. One compounding system.

* Review is the human gate.

Into the physical world

Model, action, measurement, updated model

The intelligence loop leaves the screen. Clockwork Engine models the project. Machines act on it. Sensors measure reality. Reality updates the model. The next decision starts from better evidence than the last.

Engine →Reason over a persistent project model.
Sensing →Observe the built reality.
Robotics →Act on the physical world.
3DPClose the production loop back into the model.

MODEL → ACT → MEASURE → COMPARE → UPDATE → ACT AGAIN

Work where construction data, engineering computation and physical systems meet

We work with contractors, clients, researchers and technology partners on problems where construction data, engineering computation and physical systems meet.

Discuss an engineering applicationFund or collaborate on research