Technology
The technology behind construction intelligence.
The public MVP uses prepared records to show how computer vision, multimodal AI and a construction knowledge graph could work together. Live analysis and customer data services are not connected to this preview.
Capabilities
Seven capabilities, stated with their limits.
| Capability | What it does | In Infrarity (sample project) | Limit it is designed around |
|---|---|---|---|
| Computer vision | Detects construction elements, activity and change in site imagery | Finds columns C14–C17, new since 5 Oct | Occlusion, poor light and viewpoint drift; low-confidence results go to a person |
| Multimodal AI | Reasons across images, text, drawings, video and structured data together | Links IMG-1162 to S-204 Rev.5 and STR-043 | Every answer cites its sources; no source, no claim |
| Construction knowledge graph | Stores project entities and the relationships between them | Building → Level → Zone → Element → Drawing → Activity → Contractor | Built only from the project's own data |
| Temporal intelligence | Understands how the site changes between captures | Level 4 Zone C from 51% to 63% in four days | Comparisons need matched viewpoints; mismatches are flagged, not guessed |
| Predictive analytics | Flags emerging risk from trends, sequencing and the critical path | Level 3 Zone B rated High: 8 working days past planned start, on the critical path | Early versions explain risk from visible rules and trends, not opaque forecasts |
| Retrieval-augmented AI | Retrieves the relevant drawings, specifications and records before answering | "Latest structural drawing for Zone C" returns S-204 Rev.5 | Retrieval is limited to project data the user is permitted to see |
| Spatial intelligence | Places every observation in a building, level, zone and element | Photo pinned to Level 4 Zone C, near column C17 | The MVP relies on zone tags chosen at upload; automatic localisation comes later |
Pipeline
From capture to insight.
- 01
Ingest
Photos, video, drawings and schedules arrive by upload.
- 02
Normalise
Capture time, location and zone tag are attached.
- 03
Analyse
Detection and image understanding, with confidence.
- 04
Link
Observations join zones, elements, drawings and activities.
- 05
Compare
Against the previous capture and the plan.
- 06
Score
Progress, variance and risk, by readable rules.
- 07
Explain
Copilot answers, alerts and reports, each with evidence.
Human in the loop
AI suggests. People decide.
- Confidence bands
Shown, flagged or held back
Observations at 80% confidence or higher are shown. Between 50% and 79% they are marked "Needs review". Below 50% they are kept but not shown.
- Confirmation
Nothing moves on its own
Progress figures from photos are suggestions until a project manager confirms them. Every confirmation is recorded with who, when and what changed.
- Corrections
Every fix makes it better
Corrections to labels, zones and percentages are stored against the observation and used to evaluate the models before each release.
Architecture
Product logic is independent of the cloud it runs on.
A future connected system is designed around internal service interfaces, with infrastructure adapters beneath them.
The public MVP is a hosted static site. Production hosting, data services and access controls must be chosen and verified before use with real project information.
- StorageServiceUpload and download links, object lifecycle and retention
- AIModelServiceText and multimodal model calls, the Copilot tool loop, embeddings
- VisionInferenceServiceFrame extraction now; a trained construction detector next
- DatabaseServiceRelational data, vector search, row-level security
- EventQueueBackground jobs, domain events, scheduled triggers
- AuthenticationServiceSign-up, sign-in, multi-factor authentication
- NotificationServiceEmail and in-app notifications
Explore the sample project.
Walk through the sample project from site capture to weekly report.