VisionAgent 3.0 / Presentation Brief

Can One Vision Agent Cover an Entire Factory?
Principles and Practice of Factory-Wide Vision

1 Vision AgentPowers 16 lines and 80 4K line-scan cameras (tray case)
20 StationsOne agent handles SOP error-proofing for engine assembly
1–5 Images/ClassStarting samples; on-site fine-tuning takes about 2 hours
↓ 70%Inference compute and software acquisition cost (tray project estimate)

In factories, AI vision was first used for defect detection, then assembly inspection, and now SOP monitoring. Add routine tasks such as code reading and positioning, and a natural question follows: if one system can do all of these, can it cover an entire factory’s vision needs? My presentation at VisionChina Shenzhen on October 27 explored exactly that question. This article turns the presentation into a written account and adds the calculations and project evidence that time did not allow me to cover on stage.

My answer is yes—but only when three conditions are met at the same time: the visual foundation model must deliver few-shot learning and strong generalization; the Vision Agent must be general-purpose, easy to use, and fast to deploy; and the architecture must use one-to-many edge deployment to create compelling economics. Without any one of these, factory-wide vision either cannot be built or cannot be scaled. One major challenge remains unresolved: imaging. I return to it at the end.

TALK 01Digital-presenter edition, 11 minutes 14 seconds, followed by the moderator’s remarks. The script was refined from the live presentation to remove repetition and filler while preserving all data and cases. The narration was synthesized using recordings of Dr. Cui as a voice reference; it is not a verbatim recording from the venue.
VisionChina Shenzhen seminar: a full audience and the opening slide of the factory-wide vision presentation on screen
ON SITE 01The presentation venue on October 27, 2026, with the opening slide on the main screen.

QUESTION

What Must Factory-Wide Vision Actually See?

Vision tasks in a factory fall broadly into three groups: inspecting product defects, verifying assembly status, and monitoring whether operations follow the SOP. There are also conventional tasks such as code reading and positioning. Historically, these jobs have been split across different devices and vendors: one system for appearance inspection, another for assembly error-proofing, and yet another for SOP monitoring. Every SKU change could require collecting new samples and reconfiguring the program.

Slide 2: product quality and operating processes, with defect detection, assembly inspection, and SOP monitoring distributed across stations on the same line
DIAGRAM 01Factory vision tasks: defect detection, assembly inspection, and SOP monitoring at different stations on the same line. Slide 2.

For one platform to cover all of them, the technology must answer two questions. First, is it practical to use? Can one model handle all three task types, and can a product change avoid a new sample-collection effort from scratch? Second, is it affordable? Can the economics support factory-wide rollout rather than one or two pilots? Everything that follows addresses these two questions.

PRODUCT FORMS

Two Product Forms: Edge Agent and Local Agent

Vision Agents come in two product forms. Most capabilities are the same; the difference lies in how they are integrated.

Item Edge Agent (LBZN-VA-I) Local Agent (LBZN-VA-I-F)
Deployment Centrally deployed beside the line; cameras connect over industrial Ethernet Integrated inside the machine
Best suited for Multiple lines, multiple stations, and mixed tasks Vision inspection and motion control within one machine
Controller Results are written to the line PLC through Modbus or similar interfaces Built-in PLC; no external PLC required
Connectivity Sized by camera resolution, sampling rate, concurrent tasks, and latency 16 camera channels and 31 lighting channels
Slide 3: edge agent centrally deployed beside the line on the left; local agent integrated in the machine with a built-in PLC on the right
EQUIPMENT 01Two product forms. Left: an edge agent centrally deployed beside the line and connected over an industrial network. Right: a local agent integrated into the machine with a built-in PLC. Slide 3.

Factory-wide coverage relies primarily on the edge agent, which is the focus of the sections below.

EDGE DEPLOYMENT

Why the Edge: Neither at the Endpoint nor in the Cloud

In the convolutional era, models were trained on large datasets, usually in the cloud or on dedicated training servers. A visual foundation model is pretrained and needs only a few on-site samples for adaptation, so training can be completed at the edge. Inference and training run on the same device—integrated training and inference. This is the first prerequisite for edge deployment.

The second prerequisite is networking. The edge and endpoints communicate over conventional industrial Ethernet: camera images go directly to the agent, and results return to station displays, PLCs, and alarm lights. A copper Ethernet segment can run up to 100 meters; fiber can bridge longer distances. This allows one agent to connect stations along a line and across multiple lines, regardless of overall line length.

Slide 4: edge deployment with one integrated training-and-inference device connected upward to MES and SCADA and outward to stations on multiple lines
DIAGRAM 02Edge deployment: one integrated training-and-inference device connects upward to MES/SCADA and downward to stations across multiple lines. The number of lines and stations shown is illustrative; actual device capacity is sized by resolution, sampling rate, concurrent workloads, and latency. Slide 4.

Upstream, the edge agent connects to MES or SCADA. MES sends the recipe, identifies the current SKU or model, and the agent switches accordingly. Downstream, each station has at least one camera. Operator stations add a display that identifies an incorrect step and prompts rework. Stations that divert or reject products also add a PLC and actuator. These are the endpoint resources.

EQUIPMENT 02The edge agent used in the tray project, installed in a cabinet beside the line (green indicator illuminated). About 80 cameras connect to this single unit. On-site footage, 10 seconds.

The Two Extremes of One-to-One Endpoint Deployment

Why not place compute at every endpoint? A one-to-one architecture tends toward two extremes: with one camera, most compute sits idle; with ten cameras, the device cannot keep up. GPUs remain expensive resources. Provisioning each station for its own peak demand leaves much of that capacity unused most of the time.

One-to-One at the Endpoint

  • Dedicated compute at every station, sized for peak load
  • Underused with few cameras; overloaded with many
  • Maintenance targets increase linearly with station count

One-to-Many at the Edge

  • One integrated training-and-inference device shared by all stations
  • Staggered workloads; channel count sized by aggregate throughput
  • Centralized recipes, models, and backups
Concept: stations A, B, and C each reserve their own compute window on the left; their workloads are staggered on one shared timeline on the right
DIAGRAM 03Per-station reservation versus staggered sharing: compute jobs from stations A, B, and C run at different points on one timeline. Pooling removes the idle windows each station would otherwise reserve for its own peak. Conceptual example only; not actual capacity or measured latency.
Concept: investment F in the platform and shared capabilities is spread across N stations, giving F/N per station
DIAGRAM 04Investment F in the platform and shared capabilities is amortized across N stations. This does not imply that each station’s on-site engineering cost also falls to F/N.

Capacity Must Be Calculated from Throughput

How many stations can one agent support? The answer is not a slogan; it is calculated from image dimensions, inference throughput, and cycle time. Consider the tray project (see “One-to-Many—Many More” for the full derivation). Inference uses 700×700 tiles, device throughput is estimated at 200 images per second, and the on-site acquisition cycle is 20 seconds. Each 4K line-scan camera produces the equivalent of one 4096×4096 frame per cycle, split six ways in each dimension into 36 tiles.

4,000 ImagesInference budget per cycle: 200 images/s × 20 s
2,880 ImagesActual load per cycle for 80 cameras: 80 × 36, or 72% of budget
22 LinesTheoretical limit per unit: ⌊4,000 / 36⌋ = 111 cameras, or 22 five-camera lines

The actual deployment stops at 16 lines, preserving 28% headroom for recapture and review, new-product trials, and future expansion. No matter how capable a system is, if it is too expensive it remains a pilot and cannot scale. Edge deployment offers the best economics: one agent powers many—many more.

GENERAL-PURPOSE

General-Purpose: One Foundation Model for Defects, Assembly, and SOPs

At the core of the Vision Agent is a visual inspection foundation model—a Transformer-based foundation model rather than a convolutional network. The first major difference is sample efficiency.

Convolutional models require large datasets—often thousands of images for one station. In a one-to-many deployment, the number of stations multiplies, and there may be ten or twenty SKUs. A rough calculation of 20 stations × 20 SKUs × 1,000 images yields about 400,000 images to collect and annotate, making deployment difficult and expensive. With a pretrained foundation model, each defect or target for a specific SKU can start with only 1–5 samples.

Item Small Convolutional Model Visual Inspection Foundation Model
Starting samples Thousands per station 1–5 per target class
Changing SKU Collect and train again Add a few samples and fine-tune on site
Training location Usually cloud-based or on a training server On the edge device with integrated training and inference
Task scope A dedicated model for each scenario One foundation model shared by defect detection, assembly inspection, and SOP monitoring
Output Different models for bounding-box detection, classification, and other tasks One form only: pixel-level semantic segmentation
Slide 5: a small CNN model requiring a dedicated model and extensive labeled data for each scenario on the left; a few-shot, cross-domain visual inspection foundation model on the right
DIAGRAM 05Small CNN model versus visual inspection foundation model: the former needs a dedicated model and extensive labeled data for each scenario; the latter generalizes across domains from a few samples. Slide 5.

It is not trained exclusively for one domain. It can detect defects, verify missing, incorrect, or reversed assembly, and monitor SOPs. The model has just one output form: pixel-level semantic segmentation, without separate branches for bounding-box detection. Segmentation is the most general representation: a defect is a region, a component is a contour, and an operator’s hands and tools are contours as well. All can be represented in the same way, then interpreted by event logic and state machines.

One distinction is important. Using a generative model to create defect images for a small convolutional model, or using a large language model as a machine interface, sits outside the inspection pipeline and serves only as assistance. The foundation model discussed here operates inside the inspection pipeline and produces the decision directly. Because it already needs very few samples, it does not rely on synthetic data augmentation.

EASE OF USE

Easy to Use: Drag-and-Drop Orchestration with Fewer, Broader Operators

On-site configuration requires no coding. Drag operators from the library onto the canvas, connect them, and assign the camera, model, and output target to each node—the workflow is complete. Compared with earlier drag-and-drop tools, this approach differs in two ways: the interface is clean, and the operators are highly abstracted and few in number.

Slide 6: drag-and-drop orchestration, with connected operators on the canvas and the operator library and external protocols on the right
UI 01Drag-and-drop orchestration: add operators and connect the flow, with an independent task configured for each station. The operator library and external protocols appear on the right. Slide 6.
The My Tensor Workflows page in VisionAgent, showing multiple workflow cards
UI 02“My Tensor Workflows”: multiple workflows for different stations and tasks managed side by side on one platform.

Defect detection, assembly inspection, and SOP monitoring use the same operator chain. Only the model classes, event definitions, and state machines differ. This is why one agent can run all three task types in parallel.

RAPID DELIVERY

Rapid Delivery: Start with Only 1–5 Samples

Rapid delivery also comes primarily from the foundation model. Annotation requires only painting over the target, and each class can start with 1–5 samples.

1–5 Images/ClassStarting samples
< 0.5 HourTraining with a dataset of up to 100 images
2 HoursOn-site fine-tuning when imaging and equipment are already in place
Semantic-segmentation annotation page for an engine assembly station, class “place gasket”: manually painted hand region on the left, adapted model output on the right, and 159 samples in the list
UI 03Semantic-segmentation page for the “place gasket” action at an engine assembly station. Left: manual painted annotation. Right: segmentation result after adaptation. The sample list contains 159 images.

As a result, a 3C electronics project is typically delivered in seven days and may include several SKUs. The first food-container line went live in ten days. Delivery time depends on on-site equipment and engineering scope: stations with imaging and hardware already in place are fastest, while projects requiring new optical and mechanical structures take longer.

CASE 1 · ONE-TO-SIXTEEN

Case 1: One Vision Agent for 16 Disposable-Tray Lines

CASE 01On-site footage from the disposable-tray line, 62 seconds: thermoforming, robotic pick-up, imaging at the inspection station, OK/NG decisions from the Vision Agent, and robotic sorting. The screen on the left is the live operations dashboard.

Sixteen production lines, each with five 4K line-scan cameras—80 cameras in total—all connect to one edge agent. Because 4K line-scan images are large, people often assume a foundation model will be too slow and must run in the cloud. The on-site cycle time is 20 seconds. As calculated above, 80 cameras produce 2,880 tiles per cycle, consuming 72% of the device’s throughput budget.

Cost: Replacing 16 Systems with One

Per-Line Deployment · 16 Lines

  • 16 inference systems and software licenses, about RMB 30,000 per line
  • 16 maintenance targets, each upgraded and backed up separately
  • Total acquisition cost: about RMB 480,000

Centralized Deployment · Same 16 Lines

  • 1 Vision Agent including platform license, about RMB 150,000
  • One platform, shared compute, and one maintenance target
  • About RMB 9,000 per line

RMB 480,000 versus RMB 150,000—a reduction of about 70%. This comparison includes only inference compute and software acquisition. Cameras, lighting, frames, sorting equipment, and on-site engineering are required in both architectures and are excluded.

Accuracy: Full Manual Inspection Across 18 Shifts

Metric Definition Acceptance Limit Measured
Miss Rate Defective items found by full manual inspection among products classified as acceptable by the machine ≤ 1.50% 0.39%
False-Reject Rate Machine-rejected products confirmed acceptable by manual review, divided by total shift output ≤ 5.00% 3.28%
Sample Size Item-by-item full manual inspection over 18 shifts — 74,916 Items
DATA 01Tray project acceptance results for machine No. 709, this product specification, during the August measurement period. Under the definitions in GB/T 46886-2025, the false-acceptance rate was 0.33%, the false-rejection rate 3.28%, and accuracy 96.38%. See “One-to-Many—Many More” for shift-by-shift records.
Live dashboard for disposable-container inspection: four live views and detections on the left, per-item OK and NG cards on the right, and daily totals at the top
UI 04Operations dashboard: live images and detections on the left, with per-item OK/NG cards on the right. The screen shows only the status at the moment of capture and does not represent a fixed yield for other dates or products.

CASE 2 · ONE-TO-TWENTY

Case 2: One Vision Agent for 20 Automotive Engine Assembly SOP Stations

SOP monitoring verifies whether operators follow the documented procedure. This automotive engine assembly line has 20 SOP stations, all handled by one Vision Agent. Each station verifies 5–15 steps individually. If a step is missed, audible and visual alarms trigger immediately and the display identifies the omission. The project went live in 20 days.

Viewed another way, 20 stations with 5–15 steps each means one device tracks the states of 100–300 process steps simultaneously. A state machine follows the required sequence and raises an immediate alarm if a step is skipped or missed.

CASE 02Integrated dashboard, 1 minute 40 seconds: one agent handles 20 stations, with images and status centralized on one screen; each station also has its own local display.
DIAGRAM 063D illustration of the production layout, 12-second loop. It shows relative station positions and is not live footage. One agent supports entirely different tasks along the length of the line.

CASE 3 · ONE-TO-FIVE

Case 3: One Agent for Five Stations in 3C Smart-Hardware Assembly

What if there are only a few stations? Vision Agents now come in three capacity tiers; even the smallest can support ten SOP stations. This 3C project has just five stations and two task types—assembly inspection and SOP monitoring. It uses the smallest tier and still has spare capacity.

Slide 12: 3C case installation overview, articulated camera mount, MES terminal and vision display, plus imaging and hardware parameters; the center circle marks the standard mount
DIAGRAM 07The 3C site and hardware selection. The center circle shows the standard articulated mount (ring and arrow added for this article), which can hold a camera or display and can be adjusted to match the target field of view. The image on the right shows the MES terminal and vision display in front of the operator. Slide 12.
Item Selection Notes
Camera 5 MP, global shutter Field of view and small targets determine position and focal length
Lens 12–36 mm zoom One model covers different camera distances
Pixel scale 0.06–0.23 mm/px 1 mm spans approximately 4–17 pixels
Lighting Existing line lighting No additional close-range light
Compute Centralized in a cabinet beside the line Stations only acquire images and display results

Pixel scale determines the smallest visible detail. At 0.06 mm/px, 1 mm spans about 17 pixels; at 0.23 mm/px, about 4 pixels. Centimeter-scale targets such as manuals, films, and foam occupy roughly 43–167 pixels per centimeter. For focal length, working distance, and field of view at each station, see Ryan’s “Multi-Station Assembly Inspection and Process Monitoring in the 3C Industry”.

3C assembly station: an orange articulated mount supports the camera, a display sits in front of the operator, and the operator works at the station
ON SITE 02Assembly station: an orange articulated mount supports the camera, and the display is installed in front of the operator.
Vision Agent in a cabinet beside the line, with its green front-panel indicator illuminated
EQUIPMENT 03The agent in the cabinet beside the line (green indicator illuminated). Compute for all five stations is centralized here.

To accelerate delivery, we designed a standard articulated mount. Cameras and displays use the same mount, so the installation team only adjusts position and angle on site instead of fabricating custom fixtures for every station.

What Each of the Four Station Types Checks

Station Inspection Annotated Samples
Manual insertion Insert one white and two yellow manuals in sequence About 61
PET film application Align the circular hole, then press twice About 89
Bracket and foam Assembly correctness: bracket and conductive foam both present About 36
Scan and insert into housing Recognize the work surface and advance through ten steps About 98
CASE 03Manual insertion station, 30 seconds.
CASE 04PET film application station, 33 seconds.
CASE 05Bracket and foam assembly station, 50 seconds.
CASE 06Scan-and-insert station, 40 seconds.
Five parallel tensor workflows, each consisting of an industrial camera, image preprocessing, semantic segmentation, event detection, and a state machine
UI 05The project’s five tensor workflows: industrial camera → image preprocessing → semantic segmentation → event detection → state machine. Each of the five stations has one workflow, and all run on the same agent.
Annotation page for the manual station: white-manual contour on the left and the adapted model’s segmentation result on the right
UI 06Annotation page for the manual station. Left: manually annotated “white manual.” Right: segmentation result after adaptation.

MORE APPLICATIONS

24 Production Applications: One Foundation, Three Task Types

Historically, one vision system could rarely handle different task types, and SOP monitoring was especially difficult. The presentation summarized 24 production applications: 9 automotive, 3 pharmaceutical, 6 home-appliance, 2 in 3C electronics, and 4 in other industries. By task, they include 4 defect-detection applications, 4 assembly-inspection applications, 15 SOP-monitoring applications, and 1 fertilizer-mixture ratio inspection. One Vision Agent can support all of these factory vision tasks.

Slide 14: twelve representative production applications—nine automotive and three pharmaceutical
DIAGRAM 08Representative applications I: twelve automotive and pharmaceutical cases. Slide 14.
Slide 15: twelve representative production applications—six home-appliance, two 3C, and four other cases
DIAGRAM 09Representative applications II: twelve home-appliance, 3C, and other cases. Slide 15.
Industry Application Task Further Reading
AutomotiveInjection-molded part defect detection (foam between two panels)Defect DetectionVisual Defect Inspection of Injection-Molded Parts: Imaging and Lighting
AutomotiveTransmission assembly SOP monitoringSOP Monitoring
AutomotiveFuel-tank assembly SOP monitoringSOP MonitoringAutomotive Fuel-Tank Installation Process Monitoring
AutomotiveSeat wiring assembly inspectionAssembly InspectionOnline Monitoring for Missing Automotive Seat Wiring Connections
AutomotiveBody-in-white painting SOP monitoringSOP MonitoringAutomotive Painting Process Monitoring: Body-in-White Paint Booth
AutomotiveComponent insertion SOP monitoringSOP Monitoring
AutomotiveWiring-harness fuse-box SOP monitoringSOP MonitoringNew-Energy Assembly Process Monitoring: Wiring-Harness Fuse-Box Station
AutomotiveBody wax-injection SOP monitoringSOP Monitoring
AutomotiveComponent assembly inspectionAssembly Inspection
PharmaceuticalCapsule defect detectionDefect DetectionAI Visual Inspection for Capsule Appearance
PharmaceuticalCleanroom gowning SOP monitoringSOP MonitoringPharmaceutical Industry: Gowning Recognition in a Grade B Cleanroom
PharmaceuticalOphthalmic-equipment accessory-box packing SOP monitoringSOP Monitoring
Home AppliancesMotor junction-box SOP monitoringSOP MonitoringComponent Assembly Process Monitoring: Motor Junction-Box Station
Home AppliancesProduct-kit packing SOP monitoring (missing fuse or manual)SOP Monitoring
Home AppliancesProduct-accessory packing SOP monitoringSOP Monitoring
Home AppliancesKettle-packing SOP monitoringSOP MonitoringMissing-Item Detection at a Home-Appliance Packing Station
Home AppliancesCable presence and wiring-sequence inspection for microwave and combi-steam ovensAssembly Inspection
Home AppliancesAnti-static fabric assembly inspection behind large displaysAssembly InspectionHome Appliances: Monitoring the Anti-Static Fabric Application Process
3CRouter SOP monitoringSOP Monitoring
3CElectronic-component SOP monitoringSOP MonitoringSOP Process Monitoring: Electronic-Component Test Station
OtherNonwoven hook-and-loop fastener defect detectionDefect DetectionAI Appearance Inspection for Hook-and-Loop Fastener Patches
OtherSwim-cap SOP monitoringSOP MonitoringProcess Monitoring for Swim-Cap Final Inspection
OtherFertilizer blend-ratio inspection (three-color nitrogen, phosphorus, and potassium mixture)Ratio InspectionFlatten First, Then Measure the Ratio: Vision-Agent Sampling Inspection for Fertilizer Blends
OtherInfant-formula scoop defect detection (flat and contoured surfaces)Defect DetectionSix-Side Visual Appearance Inspection for Infant-Formula Scoops

These applications come from different projects with different imaging conditions, cycle times, and accuracy requirements. “—” indicates that no public article is currently available.

The infant-formula scoop deserves special mention. It combines flat and contoured surfaces, making it particularly difficult to capture completely with a fixed top-down camera. This leads directly to the final limitation.

ABOUT LINGBANG

Lingbang Intelligence: It Started with Data Collection

Lingbang Intelligence was founded in 2016. We began accumulating industrial images in 2017 and have collected approximately one billion to date, forming the basis for training our own foundation model.

Year Milestone
2016Lingbang Intelligence founded
2017Launched our first AI vision equipment and began collecting industrial images (about one billion to date)
2022Selected by China’s Ministry of Industry and Information Technology to lead a national machine-vision R&D initiative
2023Began training the visual inspection foundation model in January
2024Released the visual inspection foundation model; launched VisionAgent 1.0 in October
2025Released VisionAgent 2.0
2026Released VisionAgent 3.0 at this forum
2027Plan to launch VisionAgent 4.0 at VisionChina Beijing or Shanghai

When training began, ChatGPT was not yet widely known in China; broad discussion started around March 2023. Lingbang is a national-level “Little Giant” enterprise specializing in advanced and innovative technologies. The company holds 62 granted patents and 20 software copyrights and participated in drafting GB/T 46886—2025, General Technical Requirements for Intelligent Inspection Equipment.

LIMITATION

The Remaining Limitation Is Imaging: From Trial-and-Error Lighting to Computational Imaging

Factory-wide vision still has a major limitation: imaging for AI-based industrial inspection. Once an image is captured, model processing is already highly general—the foundation model handles the 4K imagery in the tray project both accurately and quickly. The hard part is ensuring that the image actually reveals the defect.

Industrial imaging today is still largely empirical: mount a camera above the part, move lights around it, and see whether the defect becomes visible. The problem lies in the samples. A customer initially provides some samples and the lighting is tuned to them, but unseen defects appear after deployment and the imaging setup cannot reveal them. Requiring a frozen reference sample set is difficult because customers rarely have every defect type available. This is the dilemma of traditional imaging.

The way forward is general-purpose imaging—in other words, computational imaging. The presentation used a simple example: capture four images and computationally derive two outputs, one representing shape variation and the other reflectance variation such as stains and contamination.

Computational-imaging concept: the camera remains fixed while lights from four directions illuminate the object in sequence; highlights on raised features move with the light while stains remain dark; pixel-wise calculation produces a shape map showing raised features and a reflectance map showing stains
DIAGRAM 10Computational-imaging principle using photometric stereo. The camera remains fixed while lights from four directions illuminate the object in sequence. Pixel-wise calculation places surface relief in the shape map and stains or color differences in the reflectance map. Conceptual illustration only; it does not represent specific equipment parameters.

Its physical basis is the diffuse-reflection, or Lambertian, model: pixel intensity I = ρ·(n·l), where ρ is surface reflectance, n is the surface normal, and l is the illumination direction. Capture one image under each of several known lighting directions and every pixel yields a system of equations from which ρ and n can be solved. Three directions are theoretically sufficient; a fourth improves robustness against specular highlights and shadows. The recovered n forms the shape map, while ρ forms the reflectance map. Relief defects appear in the shape map; stains and color differences appear in the reflectance map. Imaging no longer depends on a lighting angle discovered by trial and error for one defect type—that is what “general-purpose” means.

Lingbang has not yet solved this problem well enough, so we hope to work with industry peers to close the gap. In imaging, the whole industry should move toward general-purpose methods and away from trial-and-error practice. Otherwise, too many projects will be abandoned—walk through production lines in large factories and you may find half the cameras powered off. We should stop creating failed deployments and use the foundation-model revolution to raise the ceiling of the machine-vision industry.

Visual foundation models: few-shot learning and strong generalization;
Vision Agents: general-purpose, easy to use, and fast to deploy;
Edge deployment: one-to-many—many more, with compelling economics;
Factory-wide vision: defect detection, assembly inspection, and SOP monitoring.

Pinch to zoom and drag to pan; you can also use the buttons above.