AI that works where the signal doesn't.
Small models run on local hardware — which means they run in a site office, on a laptop, without connectivity. Daily logs, snag lists, RFIs, and project paperwork, processed where the work happens.
Construction AI has to work offline. A 3–7B model runs on a laptop or a local server in the site office, so daily logs, snag lists, RFIs, and project paperwork get processed where the work happens — connectivity or not. We build on what we learned shipping Xtimo, a full-cycle construction AI product built and operated by Crit Studio.
Construction's AI problem isn't ambition — it's conditions. Sites have unreliable connectivity, mixed devices, and workflows that live in photos, voice notes, and paper. A frontier API assumes a clean network path and tolerable latency; a small model assumes nothing. A 3–7B model runs on a laptop or a local server in the site office, which makes offline-first AI a deployment detail rather than a research project: daily log structuring, snag-list capture, converting field notes into RFI drafts, parsing invoices and subcontractor paperwork in the project office.
This is the one vertical where we've built and operated the product ourselves. Xtimo — built and operated by Crit Studio — is a full-cycle AI product for construction workflows, and most of what we know about field conditions, messy source data, and what site teams will actually use comes from shipping it. That experience transfers directly: the same evaluate–fine-tune–deploy–monitor pattern, pointed at your workflows instead of ours.
Paperwork processed on site.
Construction is lightly regulated on AI but heavily contractual on data: bid pricing, subcontractor rates, and programme details are commercially sensitive, and project data usually belongs to the contractual parties — not to a software vendor. Local deployment keeps pricing and project data off third-party services entirely, and keeps workforce data handling inside your existing GDPR arrangements. We're engineers, not your compliance advisers — here, the contract is usually the constraint, and local deployment satisfies it by default.
We shipped this vertical ourselves.
Xtimo is a full-cycle AI product for construction workflows — field capture to structured records — built and operated by Crit Studio. It is not a client logo; it's our own product, run in real site conditions, and the reason this page can skip the hypotheticals.
| Published result | Source |
|---|---|
| Inference costs fell 66–80% after moving routine workloads to fine-tuned small models | Forethought, published on AWS |
| Task accuracy rose from 81% to 93% while inference costs fell 50–68% | distil labs × Knowunity, vendor-reported |
| Task-specific small models to see 3× the adoption of general-purpose LLMs by 2027 | Gartner forecast |
Does it work without internet on site?
Yes — that is the point. A 3–7B model runs on a laptop or a local server in the site office, so daily logs, snag lists, and field notes get processed with no network path at all. Connectivity becomes a sync convenience, not a dependency.
What did shipping Xtimo teach you?
Most of what we know about field conditions: unreliable connectivity, mixed devices, workflows that live in photos, voice notes, and paper — and what site teams will actually use. Xtimo is a full-cycle AI product for construction workflows, built and operated by Crit Studio, and that experience transfers directly to your workflows.
Who sees our bid pricing and project data?
Nobody outside your organization. The model runs locally, so pricing, subcontractor rates, and programme details stay off third-party services entirely — and workforce data handling stays inside your existing GDPR arrangements.
Air-gap friendly by design.
- Works fully offline
- Bid data stays yours
- Sync only to your servers
- Cloud optional — never required
See if the numbers hold on your sites.
The Crit is a teardown of your AI spend and workflows: where a small model wins, where an API is fine, and where AI shouldn't be used at all. You get the numbers either way.
If your workload doesn't clear roughly 50M tokens a month and you have no privacy constraint, a frontier API is probably fine — and we'll tell you so.