Technology, disclosed with methodology.
This is the single page where we lay out how Insightek is actually built: the capability stack behind the two products, every benchmark next to its test method, the full 12-item patent list with honest status labels, and the engineering commitments that keep those labels honest.
Six layers that hold the products up.
Every product we ship draws from the same underlying stack. No product-specific magic, no black-box claims — each layer below has a clear responsibility and a clear output contract.
Vision foundation model
A large multimodal vision model that reasons about components, text, and structure — not raw pixels. This is what gives the Visual Inspection AI Agent its "shoot, confirm, ship" workflow and the Action Compliance system its robust keypoint extraction.
- Multimodal feature fusion across RGB, text, and spatial cues
- Zero-shot component and text recognition for industrial scenes
- Survives lighting, material, and minor positional drift without recalibration
- Shared backbone across both product lines — one model, two deployments
Edge-cloud hybrid inference
A dispatcher that decides which inference workload runs at the factory edge and which runs in the cloud, based on latency budget, data sensitivity, and hardware headroom. Customers choose the split; the system honors it.
- Real-time PLC-triggered pipelines pinned to the edge
- Heavy registration and re-training jobs dispatched to cloud when policy allows
- Graceful degradation to edge-only mode on network outage
- Deterministic latency targets per station under load
Semantic OK-sample learning
Instead of training a new model per defect class, the system extracts a semantic baseline from a handful of qualified OK samples. "Good" is learned once, deviations are judged against it in the product's own feature space.
- Baseline generation from a handful of OK samples, not thousands of labelled defects
- Single-click operator confirmation — no scripting, no threshold tuning
- Changeover from hours or days to minutes
- Works with industrial cameras and standard smartphones alike
Three-level action localization
For Action Compliance, judgment combines absolute (camera frame), relative (body frame), and temporal (cycle frame) scores into a weighted confidence, then buckets the output into correct / manual review / anomaly. One brittle rule cannot drive an alert.
- Absolute: hand position within the target workspace region
- Relative: hand-to-body geometry, operator-invariant
- Temporal: action phase within run cycle and motion period
- Thresholds: ≥0.85 normal · 0.70–0.85 manual review · <0.70 anomaly
On-site data sovereignty
The default deployment for every product we ship is on-premise. Video, images, structural data, events, and logs all stay inside the customer's factory network. Cloud is offered only where the customer explicitly opts in.
- Encrypted local storage with per-task namespaces
- Role-based access control and tamper-evident audit trails
- Optional structure-only mode: stores events and metrics, never raw video
- Action Compliance ships on-premise only — by design, not by omission
Methodology-first evaluation
Every number we publish is paired with its test method and baseline. If a metric cannot be published, we show the measurement method instead. This is a shipping policy, not a marketing slogan — the MethodologyTable component throws a build error if a caveat is missing.
- Every performance claim runs through the same disclosure component
- Test method, baseline, and caveat are required — no exceptions
- Full methodology reports are issued during POC, not sprinkled in press releases
- We use dot-coded status labels everywhere — never a bare certification word without a matching badge
Every number, disclosed with its test method.
These are the same benchmark rows published on each product page, consolidated here for side-by-side reading. Nothing is edited or rounded up for this page — it is a mirror of what the product teams are willing to defend in front of an engineer.
Every number on this page is disclosed with its test method. None of the numbers are marketing theatre. Where a metric cannot be published, we show the measurement method instead.
Visual Inspection AI Agent
See the product page →Figures below come from a named customer deployment on a multi-variant control-module line under NDA. The baseline column lists exactly what the customer was running before Insightek replaced it. Every row is measured end-to-end on the real line, not in a synthetic test rig.
| Metric | Value | Test method | Baseline |
|---|---|---|---|
| Engineering debug time | −95% | End-to-end time from blank line to first OK/NG on a new product variant | Customer's prior traditional machine-vision setup on the same line |
| Single-station throughput | +50% | Takt-time measurements post-deployment compared to pre-deployment baseline | Manual visual inspection at the same station |
| Programming lines of code | 0 | Engineer timesheet during model registration (photo + confirm) | — |
| Product changeover time | Minutes | Stopwatch measurement from new sample in hand to stable OK/NG | Traditional rule-based vision: hours to days |
Figures are from a single named customer deployment under NDA. Specific numbers vary by line, product mix, and lighting conditions. A full methodology report is available on request after NDA.
Action Compliance & Bottleneck Analysis
See the product page →Action Compliance is newer, so the rows below are acceptance targets validated during prototype and pilot deployment — not marketing ceilings. The "Notes" column calls out exactly how each target is scoped (per-station, per-shift, ≥ or ≤). Real values vary with station configuration, lighting, and motion complexity.
| Metric | Target | Test method | Notes |
|---|---|---|---|
| Video capture & sampling success | ≥ 99% | Ratio of successful capture tasks over total tasks | Per-station, per-shift |
| Keypoint structural stability | ≥ 95% / ≥ 90% | Inclusion rate of valid keypoints over sampled frames | Head / shoulders / feet ≥ 95% · other joints ≥ 90% |
| Primary operator sequence tracking | ≥ 98% | Continuous trajectory frames ÷ total frames | — |
| Cycle identification accuracy | ≥ 90% | Cycle boundary annotation vs. system output on sample set | — |
| Action judgment combined accuracy | ≥ 92% | System output vs. template-based ground-truth annotations | — |
| Operator manual correction rate | ≤ 10% | Proportion of uncertain-state decisions escalated to review | Accumulated over 7 days post-launch |
| Single-unit UI response time | ≤ 3 s | End-to-end round trip on typical LAN | — |
| 10-minute video full parse time | ≤ 10 min | Complete timeline analysis wall-clock time | Reference hardware |
These are acceptance targets from prototype and pilot validation, not ceiling numbers. Real values vary with station configuration, lighting, and motion complexity. A formal test report is issued during POC.
12 patents, every one with an honest status label.
The full list across the two product lines and the shared stack. Each item uses a dot-coded status badge (Granted / Pending / Published) so the read is unambiguous — we never approximate a status with a word we cannot back. Specific filing numbers and jurisdictions are available to qualified partners under NDA.
Multimodal feature fusion for industrial vision
Semantic baseline from OK-sample learning
Zero-shot product registration workflow
Three-level action localization
Edge-cloud distributed inference scheduler
Adaptive lighting invariance in industrial CV
Temporal action cycle segmentation
Hybrid pixel-semantic matching
Real-time PLC-triggered detection pipeline
Privacy-preserving on-site inference
Structure-only mode for video analytics
Industrial camera calibration pipeline for OK-sample learning
Specific filing numbers and jurisdictions are available to qualified partners under NDA. Status labels are updated on this page when the underlying filings move between states — we do not batch updates to coincide with marketing moments.
The principles that keep the numbers above honest.
These are not aspirations — they are the shipping rules we enforce in code review and in the build pipeline.
Methodology over marketing numbers
Every performance claim on the site flows through a single MethodologyTable component. That component requires a test method, a baseline, and a caveat — if any of those are missing the build throws. It is architecturally impossible to ship a bare number.
Data sovereignty as default
Every product we ship supports on-premise deployment. Action Compliance is on-premise only by design. Cloud is an option the customer opts in to, never the default we hide behind. Images, video, and events stay inside the factory network unless the customer signs off otherwise.
Honest status labels
Patents, certifications, and compliance work all use the same dot-coded StatusBadge component with a fixed vocabulary — Granted, Pending, In progress, Published. If a label does not exist in that vocabulary, we do not make one up, and we do not use softer wording to imply a status we have not earned.
POC-first go-to-market
Every engagement begins with a paid POC scoped to a real station on a real line. We do not publish a price list because real deployments differ too much in line count, variant count, and integration scope. Methodology reports are issued during POC, not after the contract is signed.
Read the methodology. Then bring us your line.
If anything on this page raises a question your current vendor cannot answer, that is the conversation we want to have. Book a 30-minute technical call with the engineering team.