Active programmes
Six lines of investigation, all in progress.
Each card names the product the work feeds, the technological uncertainty we are working on, our approach, and where it stands today. Detailed programme records are available to funding bodies and reviewers on request.
Programme 01 · Existing productActive
Grounded conversational AI over licensed listing data
An assistant that answers buyers from a brokerage's own listings and content, on our own fine-tuned model and GPUs, without ever pulling data from a board the brokerage isn't licensed for and without sending conversations to a public model.
- Uncertainty
- How to keep a generative model's answers tied to live, licence-gated listing records while it still reads naturally and never pretends to be a person.
- Approach
- Retrieval over brokerage content vectored in Redis, a licence check enforced before retrieval, and answer sanitisation on the frontend before anything is shown.
- Status
- In production as the AI Assistant. Ongoing work on valuation explanations and hand-off to the right agent.
RealViuAI AssistantIn-house model
Programme 02 · Existing productActive
Matched multi-view virtual staging
Staging one room photographed from two to four angles so that the same furniture, materials and light appear in every view, instead of four unrelated renders.
- Uncertainty
- Generative image models have no built-in notion that four photographs are the same room. Cross-view consistency of geometry, objects and lighting is unsolved for real-estate sets.
- Approach
- An anchor photograph sets the furniture story; a shared brief is carried across views; every result keeps the original beside it for review. Camera position and architecture are declared out of scope.
- Status
- Shipped as a Furnio workflow. Ongoing experiments on consistency across wider angle changes.
Furnio.aiGenerative imaging
Programme 03 · Existing productActive
Computer vision for listing media production
A suite of eight models that work on every photograph a media agency delivers: virtual staging, 3D floor-plan generation, photo tagging, room measurements, listing descriptions, feature extraction, twilight conversion and exterior enhancement.
- Uncertainty
- Deriving floor plans and room dimensions from ordinary listing photography, without a scanner or a measured survey, at an accuracy agents can use.
- Approach
- Models trained and evaluated on real shoot sets, with output landing in the delivery portal so agencies can review and sell it as an add-on.
- Status
- In production inside Realview Hub. Measurement and floor-plan accuracy remain active research.
Realview Hub8 models
Programme 04 · Existing productActive
Lead intelligence and multi-channel automation
Turning inbound SMS, email and site chat into an assigned, qualified lead with a recorded consent trail, and following up on the brokerage's own number and domain.
- Uncertainty
- Reliably judging intent and readiness from very short, informal messages, and deciding when a human should take over, without creating false urgency or breaching CASL.
- Approach
- Conversation processing on our own GPU with retrieval from Redis, rule-based assignment the brokerage designs, and every action logged against the lead.
- Status
- In production as the Automated AI Lead Engine. Calls are the next channel under development.
RealViuLead Engine
Programme 05 · Existing productActive
Licence-aware listing data platform
Serving MLS listing data to hundreds of brokerage sites while guaranteeing that a listing only ever reaches a site whose brokerage holds the board agreement for it.
- Uncertainty
- Enforcing per-board data rights at feed, API and database level simultaneously, across a multi-region database with automated failover, without slowing search.
- Approach
- A licence gate at ingestion and an allow-list check at the API, both backed by the same database record; primary in Canada with two replicas and tested failover.
- Status
- In production. Ongoing work on replication latency and sold-data history.
RealViuData platform
Programme 06 · New projectActive
Disclosure and provenance for AI-edited property images
Making it impossible to forget that an image was virtually staged: a disclosure label that starts switched on, is composited in the browser at download, and never produces an unlabelled server copy.
- Uncertainty
- Keeping provenance attached to an edited image through download, resizing and MLS sizing, on the client side, without a second upload or a trusted server step.
- Approach
- Client-side compositing at export, originals stored as separate objects, and a written evidence standard for demonstrations.
- Status
- First version shipped in Furnio. Research continues on provenance that survives re-encoding.
Furnio.aiProvenance