1. The core problem
A layout of porcelain tiles with marble veins is not a normal image-matching problem. The computer is not just looking for similar pixels. It has to understand that every tile is a physical object, every vein is a directional structure, and every edge may either connect beautifully to another tile or break the flow completely.
Add to that the brutal reality of tile work: every tile is unique, every wall is a slightly different size, every cut piece changes the math, and once you start cutting tiles you cannot undo it. So matching has to happen before the saw runs.
2. The creative breakthrough: the Sandi Decimal System
Instead of treating each tile as a blob, Sandi encodes the four edges of each tile as a directional event stream — position along the edge, angle of the vein crossing the edge, thickness, and strength. Each event becomes a decimal number. A tile becomes a short string of decimal events. That string is the tile's fingerprint.
3. Why this matters
Two tiles whose veins continue smoothly across their shared edge will produce mirror-image edge events. A connection that looks beautiful is also mathematically scoreable.
4. The matching equation
if (
positionError <= tolerance.position &&
angleError <= tolerance.angle &&
thicknessError <= tolerance.thickness &&
strengthError <= tolerance.strength
) {
connectionValid = true;
}
The weighted score gives position and angle the most importance:
score =
0.45 * positionScore +
0.30 * angleScore +
0.15 * thicknessScore +
0.10 * strengthScore
5. The struggle of teaching this to a computer
Computer vision wants to see pixels. Sandi forces it to see structure. The hard part was making the system stop pattern-matching colours and start tracking vectors. v15→v19 was about this fight. v18 introduced cut-aware vein vectors. v30.3 restored the projection pipeline cleanly.
6. The physical workflow
- Lay out all candidate full tiles face-up.
- Photograph each tile at known scale (use a Tile Match Maker scan card).
- App extracts perimeter-band signature + inner-1-inch vein vectors.
- App proposes wall layouts that minimize sliver cuts and bad corner wraps.
- Cut tile, mount, photograph the cut piece. App records the cut fragment as its own entity, linked to its parent tile.
7. Full tile vs. cut fragment
This is the single most important distinction in v18+. A finished-wall extraction may document only the visible cut tile. It should not automatically become a new full tile identity. Perspective equality and visible fraction must support it before promoting.
8. Database-ready structure
Every tile produces a record with: ID, source image, perimeter signature, vein vector list, perspective equality score, visible fraction, parent tile (if cut), role (hero/connector/buffer/terminal/reject), and the QR-Q payload. This is what the upcoming memory layer will store.
9. Main beta goal
Match leftover cut pieces to their parent tiles, score wall layouts, and produce QR-Q payloads that travel with the tile from inventory → wall → photo record. Beta is judged on whether contractors can trust the outputs.
10. Rule: no fake certainty
Tile Match Maker never invents matches or measurements. Every output carries an evidence grade (L0 → L4). If the system is unsure, it asks. If you press it to commit without enough evidence, it refuses. This is the charter that separates this from every other vision toy.
11. The complete decision loop
capture → calibrate → project → extract veins
↓
score perimeter signature
↓
suggest matches (with confidence)
↓
show continuation line
↓
user accepts, changes, or rejects
↓
save as training data