Atlas Heritage
BY ATLAS INSIGHTS
AI for hand-woven rugs

Find the right rug. See it on your own floor.

Describe what you want in plain words. Atlas Heritage reads every rug from its photograph, shortlists three to five, and shows them at true scale in your room.

Open Envision → Open Inspire

Built around a Singapore gallery of 657 one-of-a-kind Persian, tribal and flatweave rugs (rugs.sg). Prepared for the SUTD MSc DAI-E Studio, September 2026. Studio plan below.

Rug ConsultantATLAS HERITAGE
Mainly blue, about 200 × 300 cm, calm design, under $2,000.
Three pieces fit. Sizes are within 10 cm of yours.
Qashqai, wool
203 × 297 cm
Indigo field, sparse medallion, quiet border.
Hamadan, village
196 × 305 cm
Mid blue ground, small repeating boteh.
Kilim, flatweave
200 × 290 cm
Blue bands, no pile, lightest visual weight.
Price on request: 1 of 3See in my room →
What it does

A salesperson's judgement, online

Listen, narrow the choice, show the rug in the room. Three products, one pipeline.

Rug Consultant

Plain-language request in, three to five reasoned picks out. Searches the images, not just the text.

Colour, pattern, characterread from the photograph
Size, price, availabilityhard constraints from the catalogue
One grounded reason eachcited from rug literature
Agreement with salespersonprecision@5

Right Rug, Right Room

One photo of the room. Floor plane and scale recovered, rug placed behind the furniture, fit scored.

Floor plane and scalemetric depth, or one tap-measured edge
Placed behind furnitureSAM 2 masks for occlusion
Fit scoresize, palette harmony, style
Rated by interior designers30 rooms × 3 rugs

Envision

View at Home: live web AR on your own floor from a link. No app. Shipped today.

AndroidWebXR in Chrome, ARCore floor tracking
iOSApple Quick Look, USDZ
True scalerug modelled at its real dimensions
Instagram → QR → floorenvision.atlasinsights.ai
How it fits together

One pipeline, built in order

Each stage is a product the next one depends on. The red layer is the one missing piece.

QueryPHASE 1
→
ShortlistPHASE 1
→
RoomPHASE 2
→
Video and ARPHASE 3
Catalogue657 products, 932 images, versionedLive
Rug understandingColour, pattern, motif, style from images. Learned and measured.The gap
Appraisal pipelineMotif taxonomy, knowledge graph, expert reviewPrototype
Scene parsingRF-DETR-Seg + SAM 2, furniture occlusionLive
Reference libraryRug literature indexed for grounded answersReady
Why filters fail

The first thing customers ask is the one thing the catalogue never writes down

657
one-of-a-kind rugs
932
catalogue photographs
61
product texts that mention colour
50%
of products have no description at all

Attribute named in the written catalogue

Products out of 657. Colour lives in the image.

Origin
386
Era
374
Material
289
Design
185
Colour
61
The SUTD studio

Three phases, three shipped products, one gate each

A year-long MSc DAI-E Studio. One student, one faculty mentor, one industry mentor.

Phase 1 · MVP, deployed

Rug Consultant v1

Versioned dataset, indexed reference library, vision-language attributes and embeddings, hybrid retrieval, conversational layer with expert corrections.

GATE: PRECISION@5Matches an experienced salesperson's picks on 50+ real enquiries. Sales team sign-off. p95 under 5 s.
Phase 2 · Prototype

Right Rug, Right Room v1

Floor segmentation and metric depth, scale recovery, perspective compositing with occlusion, fit scoring on size, palette and style.

GATE: 4 / 5 FROM DESIGNERSMedian realism and fit rating across 30 rooms × 3 rugs. Scale error under 5%.
Phase 3 · Feasibility and research

Video, refinement, integration

Moodboard-to-video with the rug region protected, attribute heads fine-tuned on a year of corrections, end-to-end demo and evaluation report.

GATE: BEFORE / AFTERFine-tuned beats zero-shot on the Phase 1 gate. Rug fidelity within an agreed LPIPS/SSIM threshold.
What exists today

Four codebases, four lessons

None has a learned model or a measured accuracy yet. That is the studio's job.

View at HomeShippednn-athome · envision.atlasinsights.ai · WebXR + ARCore, Quick Look, cloud occlusionLesson. Cloud segmentation is too slow for live occlusion. Perfect for a still room photo.
Appraisal and archivalPrototypenn-appraisal · inspire.atlasinsights.ai · nine-step vision + LLM pipeline, review at 0.85 confidenceLesson. Zero-shot attributes look plausible and cannot be trusted. The taxonomy and review UI are reusable.
In-gallery storytellingSprint 1 of 5nn-interactive · 8th Wall image-target WebAR on a physical rugLesson. Image tracking fails on abstract rugs. Track the floor, model the rug.
Catalogue and libraryReadySquarespace export · curated rug literature · real enquiries being loggedLesson. Small, real and messy. Foundation-model features with fine-tuned heads, not end-to-end training.
The documents

Three proposals, one product

Origin

AI-Powered Rug Finder

The customer experience brief. Became Phase 1.

Roadmap

AI Rug Consultant, SUTD Studio Proposal

Work packages, gates, data, risks, form answers. Authoritative where this page differs.

Horizon

Singapore Home Look Book

Design for the home you actually live in. Phase 2 is its first working room.

For SUTD students and faculty

What the year looks like

Two days a week across three terms. A demo on the live catalogue at every meeting from Phase 1 onward.

Track

Human-Centred AI, or Open / Cross-disciplinary

Mentor

Founder of Atlas Heritage, PhD in HCI, product owner of the AR, appraisal and catalogue systems

Cadence

Weekly in build phases, bi-weekly minimum

Data

Catalogue, images, appraisal pipeline, reference library and enquiries from day one, under NDA

Infrastructure

Databricks endpoint, Vercel, API and GPU credits, test devices, showroom and sales team

IP

Company-owned under the intern model. Student keeps publication and portfolio rights.

Back to the product ↑