1. Product Overview
VisionIQ is JoNoleecy's production-grade Computer Vision SaaS platform — a unified API surface that gives any organisation access to enterprise-level image intelligence without requiring an in-house machine learning team, GPU infrastructure, or model training expertise.
Where most computer vision tooling forces a choice between pre-built models that don't fit your domain and from-scratch training that requires months of ML engineering, VisionIQ occupies the middle ground: a curated library of pre-trained, fine-tunable foundation models served through a clean REST API, with a no-code training pipeline for custom domains and a deployment target covering cloud, edge, and on-premise from a single control plane.
The three computer vision problems this platform solves — classifying what's in an image, detecting and locating objects within it, and verifying visual quality at production line speed — are structurally similar across industries. VisionIQ packages those models, the serving infrastructure, the training pipeline, and the monitoring tooling into a single subscription product rather than rebuilding it per client.
What VisionIQ Is Not
- Not a general-purpose research tool. Production-focused, optimised for reliability and speed over research flexibility.
- Not locked to cloud-only deployment. Every capability runs on-premise or at the edge with the same API contract.
- Not a black box. Every prediction includes a confidence score, a Grad-CAM explainability heatmap, and a full immutable audit trail.
2. Three Core Capabilities
Capability 1 — Image Classification
Assign one or more labels to an entire image with confidence scores and Grad-CAM attention maps. Supports single-label, multi-label, and hierarchical taxonomies. Pre-trained on ImageNet-21K with fine-tuning pipelines for custom domains.
Capability 2 — Object Detection
Locate and classify multiple objects within an image, returning bounding boxes, class labels, and confidence scores for each instance. Built on a YOLOv9 + DETR ensemble backbone, supporting real-time video streams and batch analysis. Handles up to 300 distinct object classes per model.
Capability 3 — Visual Quality Assurance
Purpose-built for manufacturing, packaging, and production line inspection. Detects surface defects, dimensional anomalies, contamination, and assembly failures at production speed — up to 120 FPS on GPU-equipped edge hardware. Every detection includes severity classification (Critical, Major, Minor) and a confidence-weighted verdict (reject, reinspect, pass).
3. Platform Architecture
VisionIQ is a multi-tenant SaaS platform built on AWS with a horizontally scalable inference layer, tenant-isolated data plane, and global CDN edge network for low-latency image ingestion. Every layer is optimised for the two constraints that matter most in production computer vision: throughput and latency.
Multi-Tenancy and Isolation
| Layer | Isolation Mechanism | Guarantee |
|---|---|---|
| API | API key scoped to tenant workspace | No cross-tenant data access at the API layer |
| Inference | Shared model serving, tenant-scoped model versions | Custom models never shared between tenants |
| Storage | S3 prefix per tenant, KMS per-tenant encryption key | Images and predictions stored in isolated paths |
| Database | Row-level security, tenant_id on every table | No query can return another tenant's records |
| Audit | Immutable audit log per tenant, exportable via API | Full prediction history and data access log |
4. Data Model
The VisionIQ data model centres on three first-class entities: Models (what does the inference), Predictions (the output), and Datasets (training data for custom models). All are tenant-scoped and fully queryable via the API.
5. API Reference
VisionIQ exposes a single versioned REST API at https://api.visioniq.io/v1. All requests are authenticated with a bearer API key. Images are accepted as URLs, base64-encoded data, or multipart uploads.
| Endpoint | Method | Description |
|---|---|---|
/v1/classify | POST | Classify an image against a model's label taxonomy |
/v1/detect | POST | Detect and locate objects within an image |
/v1/qa/inspect | POST | Run visual quality assurance inspection |
/v1/batch | POST | Submit async batch job up to 100k images |
/v1/models | GET | List available pre-trained and custom models |
/v1/models/{id}/train | POST | Trigger custom fine-tuning job |
/v1/predictions/{id} | GET | Retrieve prediction with full audit detail |
/v1/datasets | POST/GET | Create and manage training datasets |
/v1/webhooks | POST/GET | Configure webhooks for async results |
SDK — Python Quick Start
6. Model Zoo
VisionIQ ships with 11 pre-trained models covering the most common industry domains, available immediately on any plan. Custom fine-tuned models are stored privately in your tenant workspace.
| Model ID | Type | Domain | Classes | Accuracy |
|---|---|---|---|---|
general-classify-v5 | Classification | General purpose | 1,000 | 98.4% top-1 |
retail-product-v3 | Classification | Retail / e-commerce | 2,400 | 97.8% top-1 |
medical-imaging-v2 | Classification | Medical (X-ray, skin) | 128 | 96.1% AUC |
document-type-v2 | Classification | Document classification | 85 | 99.1% top-1 |
general-detect-v4 | Detection | General purpose | 300 | 0.91 mAP |
logistics-v2 | Detection | Warehouse / logistics | 48 | 0.94 mAP |
people-safety-v3 | Detection | PPE / safety compliance | 24 | 0.96 mAP |
pcb-inspection-v4 | QA Inspection | PCB / electronics | 32 defect types | 0.3% FNR |
surface-defect-v3 | QA Inspection | Metal / plastic surfaces | 18 defect types | 0.4% FNR |
packaging-qa-v2 | QA Inspection | Label / packaging | 12 defect types | 0.5% FNR |
food-safety-v2 | QA Inspection | Food & beverage | 22 defect types | 0.6% FNR |
7. Custom Model Training
When no pre-trained model fits your domain, VisionIQ's fine-tuning pipeline lets you train a custom model on your labelled images — no ML engineering required. GPU provisioning, hyperparameter optimization, validation, and deployment are fully managed.
Upload Your Dataset
Upload labelled images via dashboard or API. Supports COCO JSON, Pascal VOC XML, and CSV annotation formats. Minimum 200 images per class recommended; the platform flags class imbalance automatically.
Select Base Model and Config
Choose a foundation model to fine-tune (ViT-L/14, EfficientNetV2, YOLOv9). Set accuracy targets and budget — hyperparameters are selected via Bayesian optimization, with manual override available.
Automatic Training on Managed GPUs
A managed NVIDIA A10G GPU cluster trains your model. Typical fine-tuning completes in 15–90 minutes. Real-time training metrics stream to your dashboard throughout.
Evaluate and Validate
Automatic evaluation on a held-out validation set produces a confusion matrix, per-class precision/recall, and example failure cases. You approve before the model becomes deployable.
Deploy as API Endpoint
One click deploys your custom model to the same API surface as all platform models. Zero-downtime blue/green deployment. Your model_id is immediately usable in any request.
8. Edge Deployment
VisionIQ Edge brings the full inference capability to on-premise hardware — factory floors, retail locations, hospital networks, or any environment where images cannot leave the building.
VisionIQ Edge runs the identical REST API on a Docker container deployable on any NVIDIA Jetson device, x86 server with GPU, or CPU-only server. Your application code doesn't change — only the endpoint URL switches from api.visioniq.io to your local device address.
| Deployment Target | Throughput | Latency | Power |
|---|---|---|---|
| NVIDIA Jetson Orin NX | Up to 45 FPS | <25ms | 10–25W |
| NVIDIA Jetson AGX Orin | Up to 120 FPS | <12ms | 15–60W |
| x86 Server + A10G GPU | Up to 280 FPS | <8ms | 150W |
| CPU-only x86 | Up to 8 FPS | <130ms | 65W |
Edge models are TensorRT-optimized at export, reducing model size by up to 4× and improving throughput 2–3× versus unoptimized PyTorch. Model updates are pushed via the cloud management plane and applied with zero-downtime rolling restarts.
9. Integrations
| System | Integration Type | Use Case |
|---|---|---|
| AWS S3 / Azure Blob | Native connector | Read images directly from object storage without re-uploading |
| Kafka / SQS | Event consumer | Stream images from a message queue for real-time inspection |
| SAP / Oracle ERP | Webhook + REST | Trigger QA verdict actions directly in your ERP |
| Snowflake / BigQuery | Prediction export | Export all prediction history for analytics |
| Slack / Teams | Alert webhook | Notify teams when defect rates exceed thresholds |
| Grafana / Datadog | Metrics export | Monitor accuracy drift, throughput, and latency |
| Label Studio | Dataset sync | Send low-confidence predictions for human labelling and active learning |
10. Industry Use Cases
Manufacturing — Automated Quality Control
Inspect 100% of production at line speed. Detect surface scratches, dimensional deviations, assembly errors, and contamination — replacing manual spot-check sampling that catches only a fraction of defects.
Retail — Shelf Compliance and Inventory
Verify correct product placement, labelling, and facings at every camera-equipped fixture, continuously. Detect out-of-stocks and planogram violations in real time.
Logistics — Package and Label Inspection
Read barcodes, QR codes, and shipping labels at conveyor speed. Flag damaged packaging before dispatch. Automate sortation decisions based on visual inspection output.
Healthcare — Medical Imaging Triage
Pre-screen radiology images to prioritise urgent cases for clinician review. Classify skin conditions, flag anomalies in pathology slides — always with a human clinician making final decisions.
Safety — PPE and Access Compliance
Verify hard hat, high-visibility vest, and safety boot compliance across production areas. Alert supervisors in real time when PPE violations are detected.
11. Pricing
Full API access, all 11 pre-trained models, up to 10,000 free predictions. Custom model training available on Growth and above.
| Plan | Predictions / Month | Custom Models | Edge Deployment | Price |
|---|---|---|---|---|
| Starter | 50,000 | — | — | $199 / mo |
| Growth | 500,000 | Up to 5 | — | $999 / mo |
| Scale | 5,000,000 | Unlimited | Up to 10 devices | $3,499 / mo |
| Enterprise | Unlimited | Unlimited | Unlimited | Contact sales |
Batch processing is billed at 0.7× the standard per-prediction rate. Edge device licences on Scale and Enterprise are included in the subscription. All plans include 99.95% API uptime SLA and full audit log export.