Nyckel
Nyckel is a Machine Learning API that provides developers with lightning-fast machine learning capabilities for computer vision, natural la...
Last verified:
What is Nyckel?
Nyckel is an AI platform that enables users to build highly accurate machine learning models for image, text, tabular classification, search, and detection without needing ML expertise or a PhD. Users upload labeled examples, and Nyckel automatically tests hundreds of models to find the best one, achieving reliable predictions with consistent outputs, confidence scores, and fast inference times. It's designed for applications requiring repeated decisions like spam detection, fraud detection, content moderation, and defect identification, contrasting with generative AI by focusing on discriminative tasks trained on custom data.
Key features include rapid model training in minutes, active learning to suggest improvements and reduce annotation time, fully hosted infrastructure handling scaling, GPUs, and maintenance, plus invoke capture for ongoing model refinement. It supports text, image, and tabular data inputs, with RESTful API endpoints for easy integration. Security is prioritized with no data sharing, SOC2 certification, and HIPAA compliance.
Nyckel is ideal for developers, product teams, and businesses automating tagging, categorization, or moderation tasks, such as companies like Gardyn and Gust using it for image sorting and content labeling. It's perfect for non-experts wanting production-ready ML APIs without infrastructure hassles, enabling quick iteration on custom classifiers for specific domains like product photos, X-rays, or customer support tickets.
The platform uses cross-validation for accuracy estimation, provides class-level metrics, and recommends balanced, representative training data. With features like label metadata, external IDs for tracking, and enterprise options for massive scale, Nyckel streamlines ML deployment from prototype to billions of invokes.
Nyckel pricing
Pricing model: Freemium
Development: Free, max 1,000 invokes/month, up to 10 functions/account, classification 5,000 samples/20 labels, search 5,000 samples, detection 50 samples; first month free, free model training. Production: $50 min spend/month, 10,000 invokes included ($0.005 additional), up to 1,000 functions, classification 10,000 samples/50 labels, search 1M samples, detection 100 samples. Growth: $500 min spend/month, 500,000 invokes included ($0.001 additional), up to 10,000 functions, classification 25,000 samples/200 labels, search 10M samples, detection 500 samples. Enterprise: Custom, scales to thousands of labels, millions of samples, billions of invokes; includes SSO, VPC deploy, model export, SLAs, dedicated support.
Nyckel pros
- No ML expertise required
- Models train in minutes
- Tests hundreds of models automatically
- Active learning for quick improvements
- Fully hosted with scaling included
- Millisecond inference speeds
- Consistent predictions every time
- Confidence scores provided
- Supports image classification
- Supports text classification
- Supports tabular data
- Invoke capture for refinement
- SOC2 and HIPAA compliant
- No data sharing policy
- Free development tier available
- Easy RESTful API integration
- Cross-validation accuracy estimates
- Custom labels and metadata
Nyckel cons
- Free tier limited to 1,000 invokes/month
- Max 10 functions in development plan
- Detection limited to 50 samples free
- 25 RPS and 25 concurrent request limits
- Tokens expire after one hour
- Text context limited to 512 tokens
- Production requires minimum spend
- Enterprise features cost extra
- No model export in lower tiers
- Samples must be balanced for best results
Frequently asked questions about Nyckel
What data types does Nyckel support?
Nyckel supports Text, Image, and Tabular inputs. Tabular includes Text, Number, and Image fields. Images can be URLs, data URIs, or binary multipart uploads; text as JSON strings; tabular as JSON objects with field keys.
How does model training work?
Upload labeled samples, Nyckel automatically tests hundreds of ML models via cross-validation to optimize accuracy. It trains on all data for the final model, shows overall and class-level accuracy, and uses active learning to suggest improvements.
What is invoke capture?
Invoke capture saves informative production invokes as samples for annotation, enabled by default (query ?capture=true). Use externalId for tracking; it helps refine models with real-world data without manual collection.
How is authentication handled?
Use JWT access tokens from /connect/token endpoint with client_id and client_secret. Tokens expire in 1 hour; include as Authorization: Bearer header. Secrets found in account Security section.
What are the API throttling limits?
25 requests per second (RPS) and 25 concurrent requests. Exceeding returns 429 error. Enterprise plans can relax these limits.
Is my data secure on Nyckel?
Yes, no data or models shared with others. SOC2 certified for security/privacy, HIPAA compliant for patient data. All processing follows compliance standards.
What accuracy metrics are provided?
Cross-validation accuracy: correct predictions/total samples overall and per class. Final model trained on all data; bars show class performance in console.
How to handle long text inputs?
LLMs have 512 token context (~300-500 words). Use full text (first part often suffices) or preprocess by splitting into shorter samples for better performance.
What is the difference between classification types?
Classification for labeling images/text/tabular; Search for similarity ranking; Detection for object localization in images. Limits vary by plan, e.g., detection has lower free sample caps.
Can I get multiple label predictions?
Yes, use ?labelCount=N query param in invoke for top N labels with confidences. Default is highest confidence only; includeMetadata for label details.