Insightek.ai
Technology

One platform. Three layers. Verifiable metrics.

All three product lines — PCB micro-hole metrology, post-dicing edge inspection, and semiconductor test-pin arrays — share one technology base: high-resolution imaging with high-speed stitching, AI compute with dedicated models, and equipment with factory integration. Every key metric comes with a definition and a calibration method, not just a number.

Platform

One tech base, three layers — shared by every product line.

All three product lines draw from the same stack. No product-specific black box: each layer has a clear job and a clear output, from raw imaging through to a verdict your line can act on.

01

Layer 1 — Imaging & stitching

High-resolution cameras, multi-mode illumination, and a motion stage feed high-speed, micron-level image stitching — delivering full-panel, full-wafer, or full-array coverage while keeping sub-field resolution.

  • High-resolution cameras + multi-mode illumination + motion stage
  • High-speed, micron-level image stitching
  • Full-panel / full-wafer / full-array coverage
  • Sub-field resolution retained across the stitch
02

Layer 2 — AI compute & models

A hardware-accelerated platform loads gigapixel images in seconds and runs parallel inference. Transformer-based domain models perform metrology and real-vs-false defect discrimination in the same pass.

  • Gigapixel images loaded in seconds
  • Parallel inference on a hardware-accelerated platform
  • Transformer-based domain models
  • Metrology and real-vs-false discrimination in one pass
03

Layer 3 — Equipment & integration

Full-machine integration runs from image acquisition and motion control through to verdict output, with PLC, MES, SPC, and SECS/GEM interfaces and report / trend data output.

  • Acquisition → motion control → verdict output
  • PLC / MES / SPC / SECS-GEM interfaces
  • Report & trend data output
  • Deployment line-side, inspection station, repair station, or lab
Metrics

Every number comes with its definition and calibration method.

These metrics are platform-wide and shared across all three product lines. For each one we publish what it means, how it is measured, and how it is accepted on your samples — the bare number is never the deliverable.

Every published metric is explainable, calibratable, and verifiable — we provide the method, not just the number.

Metric definitions — how to read every number

See the product page →

The four metrics that decide whether an inspection is trustworthy. Each is defined, tied to how it is measured, and tied to how it is accepted on your samples during adoption.

Platform metrics — definition and acceptance method
Metric Typical Definition & how it is measured Acceptance
Measurement repeatability 1–3 μm (±3σ) Dispersion (±3σ) of repeated measurements at the same position on the same sample; varies with objective magnification, pixel size, and sample surface condition. GR&R / repeatability calibrated on your actual samples at adoption and written into the technical agreement.
Detection capability Set on samples Smallest defect size that can be stably detected; depends on optical resolution and defect contrast; calibrated with your defect and borderline samples. Detection-limit, overkill, and escape criteria fixed on a mutually confirmed sample set.
Inspection speed < 0.5 s / cycle Baseline for a single compute configuration; varies with field size, resolution, and compute; compute is scalable. Configuration derived from your target takt, confirmed by measurement during pilot production.
Verdict consistency Blind-tested Repeat-consistency of AI verdicts on the same batch; models improve continuously as new samples are added. Blind-test sample set compared against human re-judgment.

Every published metric is explainable, calibratable, and verifiable — we provide the method, not just the number. Typical values vary with optical configuration, field size, and sample condition.

Headline platform figures (typical)

See the product page →

Typical values for a representative optical configuration. They move with field size, resolution, and sample condition, so we treat them as a starting point to confirm on your line — not a fixed promise.

Typical platform figures — confirmed on your line
Capability Typical What drives it Note
Field size 250–1000 mm High-resolution cameras and multi-mode illumination on a motion stage, joined by high-speed stitching. Varies with optical configuration.
Measurement repeatability 1–3 μm (±3σ) Repeated measurement at the same position on the same sample. Calibrated on your samples; GR&R written into the technical agreement.
Targets per field of view 10M-class Sub-field resolution maintained across the stitched field. Varies with field size and resolution.
Stitch + inspect per cycle < 0.5 s Gigapixel load and parallel inference on one compute configuration. Compute is scalable to your takt.

Typical values only; they vary with optical configuration, field size, and sample condition, and are confirmed by measurement during pilot production. Every metric here is explainable, calibratable, and verifiable.

In-house R&D

Models, software, compute, and equipment — all developed in-house.

The models, the software platform, the AI compute unit, and the standard equipment are all self-developed. One technology base carries all three product lines — which is exactly what lets us define and calibrate every metric rather than quote a number from a black box. The badge on each block marks which of the three platform layers it belongs to.

All in-house
Imaging & stitching AI compute & models Equipment & integration
01 Imaging & stitching

High-resolution cameras & multi-mode illumination

Category Shared platform
02 Imaging & stitching

High-speed, micron-level image stitching

Category Shared platform
03 AI compute & models

Hardware-accelerated AI compute unit

Category Shared platform
04 AI compute & models

Transformer-based domain models (metrology + real-vs-false)

Category Shared platform
05 Equipment & integration

Full-machine integration & motion control

Category Shared platform
06 Equipment & integration

Standard equipment & line interfaces (PLC / MES / SPC / SECS-GEM)

Category Shared platform

The models, software platform, AI compute unit, and standard equipment are all self-developed. Adoption is phased — start with software, move to an integrated unit, then standard equipment. Actual capability is always subject to on-site validation on your samples.

How we work

The method, not just the number.

These are the commitments behind every figure above — the reason our numbers can be checked rather than taken on faith.

01

Explainable, calibratable, verifiable

Every published metric comes with its definition and calibration method. We hand over the method, and we write GR&R and acceptance criteria into the technical agreement — not just a number in a datasheet.

02

Accepted on your samples

Detection limit, overkill, escape, repeatability, and verdict consistency are all calibrated on your good, defect, and borderline samples — and against a blind-test set re-judged by your team — then confirmed during pilot production.

03

Data you can act on and trace

OK/NG verdicts, coordinates, size deviations, and defect classes; panel / wafer / array maps and heatmaps; traceable re-judgment; reports and trends exported as CSV, images, or your system's format.

04

Fits your line

PLC, MES, SPC, and SECS/GEM interfaces, with deployment line-side, at an inspection station, at a repair station, or in the lab. Facility requirements follow the model selection sheet for the configuration you choose.

Bring your samples. We will show you the method.

Every metric on this page is explainable, calibratable, and verifiable on your samples. The first step is a sample and process review — its deliverable is a feasibility conclusion, not a sales quote.