logo asset

Services / Service line 04

AI Engineering for AEC

AI made prototypes cheap. We make them ship.

Approach
Deterministic-first
Entry point
Readiness Audit
Engineering
Python · C#/.NET

02problem

The problem

Building a convincing AI demo for a construction workflow now takes an afternoon. Building something a hundred people rely on for a document that gets sealed takes considerably longer, and the difference is not model quality.

The difference is: what happens when the model is confidently wrong on a code check; whether the same input produces the same answer twice; what it costs per document at production volume; how long a user waits inside a CAD session; and how you know — with evidence rather than impressions — that the system is working at all.

We are AI-first in how we build our own software, and we are deliberate about where AI belongs in a client’s. Compliance logic that has to be reproducible and auditable gets a deterministic rule engine. AI goes where it genuinely outperforms rules: unstructured text, drawing comparison, imagery, and assistance rather than authority.

03capabilities

What we build

Rule engines and deterministic logic

  • Configurable rule engines for code and standards compliance
  • Spatial constraint and layout validation
  • Firm-standard and project-completeness checking
  • Structured, auditable output with rule-level traceability

LLM workflows

  • Extraction and structuring of technical requirements from specifications and correspondence
  • Document and submittal review assistance
  • Natural-language interfaces over model and project data
  • Retrieval pipelines over technical documentation

Drawing and geometry intelligence

  • DWG and PDF drawing comparison and change detection
  • Geometry analysis and classification
  • Point cloud classification and scan-to-BIM assistance
  • Generative layout and configuration systems

Computer vision for construction

  • Progress detection and site imagery analysis
  • Model-to-reality comparison workflows

Production infrastructure

  • Evaluation harnesses and regression test sets for AI behaviour
  • Guardrails, confidence thresholds and human-in-the-loop review points
  • Token and inference cost control
  • Data governance: what leaves your environment, and what never does

04where_it_breaks

Where it breaks

Six things that separate an AI demo from a system people sign off on.

01 / 06

Confident wrongness in a regulated deliverable

A drawing gets sealed by a licensed professional. A system that produces plausible compliance findings with no traceability transfers risk onto that signature. Anything touching code compliance is built deterministic-first, with AI in an assisting role and every finding traceable to a rule.

02 / 06

Reproducibility

If the same model checked twice gives two different answers, the output cannot be used as documentation. Deciding which parts of a system must be deterministic is an architectural decision made at the start, not a tuning exercise at the end.

03 / 06

Evaluation

Without a labelled test set, nobody can tell whether a change improved the system or quietly broke it. Most prototypes have no evaluation harness, which is why they cannot be safely modified after the demo.

04 / 06

Unit economics

A workflow that costs a few cents in testing can cost hundreds of dollars per project at production volume. Cost per document belongs in the architecture discussion, not the invoice.

05 / 06

Latency inside a CAD session

A user with a model open will not wait forty seconds for a ribbon command. Some things belong in a background service; some belong in the desktop; getting that boundary wrong makes a technically correct system unusable.

06 / 06

Data governance

What client data goes to a third-party model provider, under which agreement, with what retention, is frequently the question that stops a pilot from becoming a deployment. It is better asked in week one.

05stack

Stack and coverage

Models and orchestration
Commercial LLM APIs with provider-agnostic abstraction, retrieval pipelines, structured output enforcement
Engineering
Python, C#/.NET, Node.js, containerised services on Azure and AWS
Domain integration
Revit API, DWG processing via ACadSharp and netDxf, IFC, APS/Forge, point cloud toolchains
Discipline
Evaluation harnesses, regression sets, cost instrumentation, human review checkpoints

06proof

Proof

In our own products

We use AI throughout our own development and ship AI-assisted features in our own products.

In client work

Our engagements in this line combine deterministic rule engines with AI assistance rather than replacing one with the other.

Case — pending

Format: the client had X. It broke at Y. We built Z. The measurable outcome was W. Two to three per page, anonymised where required.

07engagement

How we work

Most engagements in this line start with a Production Readiness Audit: bring the prototype your team or your vendor already built, and receive a written assessment of what it takes to put it in production — architecture, determinism boundaries, evaluation strategy, cost per unit of work, and the risks worth stopping for.

From there: paid discovery, fixed-scope phased delivery, then a retainer for model updates and ongoing evaluation.

We do not offer free development or free proofs of concept.

08faq

Frequently asked

We already have a prototype. What would you actually do?

Assess it against production conditions and rebuild what needs rebuilding. Usually the model interaction survives and everything around it — determinism, evaluation, error handling, cost control, integration — is the real work.

Will you use AI for code compliance checking?

Not as the authority. Compliance logic is deterministic and auditable; AI assists with interpretation, extraction and drafting, and every finding traces back to an explicit rule.

Where does our data go?

That is decided with you and written into the contract before any development begins, including provider choice, retention and whether anything leaves your environment at all.

How do we know it works?

An evaluation harness with a labelled test set, reported as measurable accuracy on your data rather than as a demonstration.

From prototype to production

AI made prototypes cheap. We make them ship.

A Feasibility Review is a 30-minute technical session with the engineers who would build it, followed by a written verdict: what the API actually allows, where it will break at scale, and what production costs. No obligation to proceed.

  • Straight answers
  • Written verdict
  • Senior engineers, not account managers