Every action on your line, auditable in real time.
An algorithm you can inspect. KPIs you can write into the contract.
A three-level scoring formula your process engineers can audit and tune. Nine acceptance targets with published test methods. Video that never leaves your factory. Three weeks to live detection — including your engineer signing off every alert rule.
Hand-tracking skeleton overlay on a worker's hands during an assembly operation, with motion cycle markers
Built for the contract conversation, not just the demo.
Each pillar answers a question your procurement team will ask.
An algorithm you can audit
Three-level scoring — absolute position, body-relative geometry, and temporal phase — combined with a published confidence formula. Your process engineers can inspect every threshold, tune every weight, and sign off before the system goes live.
Nine KPIs you can write into the contract
Each acceptance target comes with a defined test method and a measurement protocol. If we miss the number, the contract says so. No marketing claims — verification criteria.
Video never leaves your factory
Zero cloud path exists for action data. Edge inference, local storage, encrypted at rest. Network outages do not affect real-time detection. This is not an option — it is the only deployment model.
Measured at a precision electronics assembly line. Week 1: AI learns from experienced operators. Week 2: YOUR process engineer reviews every alert rule, adjusts thresholds, and confirms before go-live. Week 3+: stable detection with ≤ 10% manual correction rate.
Measurement method: Wall-clock time of the three-stage rollout. Acceptance gate: ≤ 10% manual correction rate over 7 consecutive days. The 3 weeks includes engineer confirmation — no alert ships without your team's approval.
We have named deployments and positive customer feedback — but every testimonial on this site waits for written approval from the quoted party. We do not publish fabricated or paraphrased quotes.
The most important knowledge on your line is not written down.
Tacit assembly technique lives in operators' hands. When it goes wrong, no one can see exactly where or why.
Operations depend on tacit experience.
Critical assembly knowledge lives in operators' heads and muscle memory, not in documents. New hires take weeks to stabilize.
The process is invisible.
When a defect escapes the line, it is extremely expensive to pinpoint which station, step, or operator caused it — and when.
Efficiency judgment is instinct, not data.
Takt time problems and station bottlenecks are diagnosed by gut feeling. Improvements rarely have a measurable before/after.
Manual patrol inspection is expensive and inconsistent.
Roving quality checks cost real FTE hours, catch only a fraction of misassemblies, and cannot produce a traceable audit trail.
Three stages. Three weeks. A working real-time system.
The system adapts to the line — not the other way around. No fixture redesign, no line stoppage, no written SOP.
- Week 1
Learning mode
Overhead cameras record normal production cycles by experienced operators. The system autonomously extracts action sequences, motion characteristics, and cycle anchor points.
- Week 2
Your engineer signs off
This is the week your process engineer reviews every alert rule, adjusts confidence thresholds for your specific line, and confirms before the system goes live. No alert ships without your team's approval. Individual operator variation is absorbed during this phase.
- Week 3+
Detection mode
Stable real-time operation. Sound/light alerts catch compliance violations before the part moves to the next station. Every anomaly is logged with station, step, and timestamp.
Three-stage deployment timeline: learning → correction → detection, with brand gradient
Cameras only. No fixtures redesigned, no tooling touched, no line stoppage required.
An auditable algorithm, not a black box.
S_total = w_abs · S_abs + w_rel · S_rel + w_time · S_time — three positioning scores, weighted and combined into a single confidence value. Your process engineers can inspect every component of this formula, tune the weights per workstation, and see exactly why each decision was made.
Human pose estimation
Head, shoulders, elbows, wrists, hips, knees, ankles — all tracked across frames with stability targets of 95%+ on core structural keypoints.
- 13-keypoint full-body structural extraction
- Body center-of-mass and skeletal landmark computation
- Primary operator identification across multi-person scenes
Hand tracking
Continuous hand trajectory tracking with sequence targets of 98%+, including workspace position and bilateral hand relationships.
- Hand-to-head and hand-to-body spatial relationships
- Bilateral hand separation relative to body center
- Trajectory continuity across frame transitions
Three-level action positioning
Absolute (camera-frame), relative (body-relative), and temporal (cycle-relative) scores are combined into a weighted confidence and bucketed into correct / uncertain / anomaly.
- Absolute: hand position within the target workspace region
- Relative: hand-to-body geometry, operator-invariant
- Temporal: action phase within the run cycle and motion period
- Thresholds: ≥0.85 normal · 0.70–0.85 manual review · <0.70 anomaly
Six answers your line doesn't have today.
Quality and efficiency in the same data loop. One deployment, two problem spaces solved.
SOP compliance — without writing an SOP
The system learns the standard flow from production, not from documents. Alerts fire when a step is skipped, reordered, or mis-executed — before the part leaves the station.
The exact bottleneck, not the suspected one
Cycle time, queue waits, upstream starvation, and downstream blocking — combined and auto-labeled with impact level. Data replaces instinct.
Cycle time with step-level resolution
Per-operator, per-station time distribution and variation range. Pinpoint exactly which step in the cycle causes drift.
Station busy-degree as ground truth
Effective working time vs idle/wait time — per station, per shift. The fact base your line-balancing decisions have been missing.
Shift reports that write themselves
Automated, per-shift, per-station, per-product type. Slice by operator for training focus, by product for process tuning.
Every anomaly is traceable
Station, step, timestamp, confidence score — every flagged event has a complete audit record. Shift debriefs start from evidence.
Five structural decisions behind the product.
No written SOP required
The system converts tacit operator know-how into an executable action model directly from normal production.
Quality and efficiency on one data loop
The same data pipeline outputs defect root cause AND cycle-time distribution. One deployment, two problem spaces.
Camera-only, non-invasive
No fixture redesign, no tooling changes, no line stoppage. Can be deployed during a short maintenance window.
Scales across stations by template reuse
Action templates are sedimented per workstation and product type, so new stations inherit most of the work.
Data stays in the factory
On-premise only. Video, structural data, and events all stay inside the factory network.
Nine numbers you can hold us to — each with its test method.
These are not marketing claims. They are acceptance criteria we submit to during POC. If we miss them, the contract says so. A formal test report is issued for every deployment.
| 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.
Three-week rollout at a precision electronics line.
Explicit specs. No hidden assumptions.
Two reference configurations. Production systems scale up the number of stations and parallel video streams linearly.
Single-station prototype
- CPU
- 8+ cores
- RAM
- 32 GB
- GPU
- 1× 16 GB discrete GPU
- Storage
- 1 TB NVMe SSD
Multi-station cluster
- CPU
- 16+ cores
- RAM
- 64 GB
- GPU
- 1–2× 16 GB discrete GPU
- Storage
- 4 TB NVMe SSD
- Camera: overhead, fixed, framing the worker's hands and assembly area. Minimum 1920×1080; 2560×1440 used in prototype validation.
- Sampling frequency: 2 fps baseline — higher frame rates supported for fast-motion tasks.
- No fixture redesign, no work-surface modification, no tool integration required.
Seven layers, one clean data path.
Each layer has a single responsibility and a well-defined output. Integration and extensibility happen at the display/interface layer, not in the middle of the stack.
- 01Video inputRole Camera capture, timestamp syncOutput Raw video frames + metadata
- 02Sample managementRole Frame extraction, task logging, time indexOutput Frame images + sample indices
- 03Structural layerRole Keypoint extraction, calibration, confidence scoringOutput Keypoint positions + confidence
- 04Sequence trackingRole Cross-frame skeleton tracking, primary operator isolationOutput Operator trajectory + temporal signature
- 05Cycle analysisRole Run phase segmentation, cycle identification, anchor locationOutput Cycles, phases, position indices
- 06Decision layerRole Absolute / relative / temporal combined judgment and alert rulesOutput Decisions, confidence, review states
- 07Display & interfaceRole Frontend UI, API services, MES/ANDON/QMS integrationOutput Web pages, query results, external system payloads
On-premise only, by design.
Local / factory network
On-premise onlyThis product ships only as an on-premise deployment. Video and structural data stay inside your factory network, meeting strict compliance and confidentiality requirements.
- Edge inference at the factory edge point
- Real-time alerts survive network outages
- Structured data stays inside the perimeter
Cloud is intentionally not offered for this product. Action data from the factory floor is sensitive and must stay local.
Action data is sensitive. We treat it that way.
Encrypted local storage
Key frames and structural results are stored encrypted, namespaced per task, inside the factory network.
Role-based access & full audit
Every access, modification, deletion, and personnel change is logged with tamper-evident audit trails.
Data minimization
Only the data needed for action recognition is processed. Non-critical regions can be masked and anonymized.
Structure-only mode
Optional mode that stores only events and metrics, not raw video — further reducing compliance burden.
Four commitments, on by default.
On-premise only
No cloud path exists. Everything stays inside your factory network, period.
Encrypted storage
Key frames and results encrypted at rest, namespaced per task, with tamper-evident audit.
RBAC & audit trail
Every access, modification, and deletion logged in a role-based access control system.
Data minimization
Optional structure-only mode: events and metrics only, no raw video. Compliance burden near zero.
Six packages you take full ownership of after POC.
Software package
Sampling service, analysis service, interface service, and frontend pages deployed to your servers.
Model & parameter package
Structural model, template files, weight configurations, workstation marking parameters.
Configuration templates
Per-station area configurations, sampling configurations, and parameter configurations ready to clone.
Documentation package
Operations manual, API reference, integration manual, and data specifications.
Validation package
Test logs, acceptance reports, open issue list, and version checklist.
Training package
Field training materials, usage guides, and troubleshooting documentation.
Questions your procurement team will ask.
Can I audit how the system makes decisions?
What accuracy can I write into the contract?
Does the video leave my network?
Is three weeks realistic or a marketing number?
What do I own after the POC?
See the algorithm. Review the test methods. Meet the team.
Book a 30-minute demo. We will walk through the three-level scoring formula, the nine acceptance KPIs, and a realistic three-week POC plan for your line.