Date: February 25, 2026

We hit the next milestone in our AI + computer-vision pipeline: Alpha 2.
The plate outline is now much closer to the true border - including on tilted images, where skew and perspective used to be hard.
We also detect the wells area correctly in most capture conditions.

This gives us a stronger base for what’s next: grid-accurate analysis and reliable on-device performance.

What’s new in Alpha 2

How to read the demo images

• Plate outline: magenta quad
• A1 corner: yellow marker with A1 label
• Wells area boundary: four lines (canonical orientation - the plate is shown upright, with column 1 on the left and row A at the top)
  • Left: cyan
  • Top: orange
  • Right: deep purple
  • Bottom: white

1) Plate outline: closer to ground truth

The outline is tighter. It’s more stable across angles and lighting.
That means less drift - and better alignment for downstream steps.

2) Wells area: reliably identified

We now isolate the wells area consistently.
That’s a key prerequisite for accurate grid fitting and well-level analysis.

3) Known edge cases (still to fix)

A few patterns remain:

• Wells boundary offset: sometimes the wells boundary shifts by about the wells rim thickness. In two extreme cases, the left boundary lands on printed text.
• Plate edge ambiguity: on some sides, the outline snaps to the upper edge of the plate; on others, it snaps to the outer edge. Both can look plausible, but they change the geometry slightly.
• Shadows: under strong shadows, the outline can partially follow the shadow edge instead of the true plate edge.

These are the main targets for Beta hardening.

4) Hybrid pipeline: the right approach

This milestone validates our direction.
AI handles real-world variability. Classic computer-vision tightens geometry and stabilizes results.
Together, they’re more accurate - and more predictable.

What’s working today

• Plate outline detection across typical angles and lighting
• Correct orientation handling (A1 corner marker supports consistent alignment)
• Wells-area detection is solid enough to move forward with grid work

What’s next (Beta target: June 2026)

1) More data, better training

• Expand coverage across more angles and lighting
• Add more edge-case samples (especially hard cases)

2) Stronger edge-case robustness

• Better handling of reflections and low contrast
• Better handling of strong shadows (follow the plate, not the shadow)
• Reduce boundary drift so the wells boundary lands consistently on the wells edge (not rim thickness or text)

3) Smartphone integration

• Integrate detection into the current Spotxel® Reader app
• Optimize speed and memory for on-device use
• Apply packaging/quantization for reliable mobile inference

Release timing update

In our October 31, 2025 update, we expected first public builds around January 2026.
After broader testing - especially tilted images and tough lighting (reflections, low contrast, shadows) - we shipped Alpha 2 internally to meet our quality bar.

Next milestone: Beta in June 2026, pending final edge-case validation and mobile performance.