Nyckel

Nyckel offers a comprehensive platform for developing and deploying machine learning models for image and text classification, search, and detection. Our technical infrastructure handles all aspects of model deployment, from server management to GPU optimization, while our active learning capabilities enable rapid development of accurate custom classifiers. The platform provides both pre-trained models for specific tasks and tools for building custom solutions, with robust security measures and compliance certifications ensuring trusted data processing.

Company Overview

Nyckel specializes in building accurate machine learning models for image and text classification, search, and detection. Founded by ML engineering leaders from top AI innovators, the company offers both custom classification tools and pre-trained models.

Technical Infrastructure

Nyckel handles server costs, bug fixing, scaling, and GPU optimization through their hosted infrastructure. They maintain robust security measures, including SOC2 certification and HIPAA compliance, with strict policies against internal data sharing.

Custom Classifier Development

Developers can create custom image classifiers using Nyckel's platform, which automatically tests hundreds of models against user data. The process requires just a URL for brand logos and allows for quick iteration on model development.

Pre-trained Models

The platform offers several pre-trained models including:

• Product Colors: Analyzes image content for eCommerce applications

• Blurry Images: Evaluates image sharpness for content moderation

• Logo Identifier: Detects specific logos across various images

API Functionality

The platform provides RESTful API endpoints for managing functions, samples, and labels. Supported programming languages include Python, JavaScript, and PHP. Authentication requires a JWT token issued by Nyckel's identity provider.

Data Handling

Nyckel relies on cross-validation to estimate function accuracy, training multiple models on most but not all of the samples. Each model predicts labels for the remaining samples, providing a comprehensive accuracy assessment.

Technical Platform

Developers can create custom image classifiers using Nyckel's platform in just minutes. The process involves minimal technical requirements, allowing users to build accurate models without a PhD. The platform automatically tests hundreds of models against user data, providing an efficient way to develop reliable classification systems.

Key features include active learning capabilities that help improve model accuracy and hosted infrastructure that manages server costs, bug fixing, scaling, and GPU optimization. This comprehensive approach allows developers to focus on their core applications while leveraging Nyckel's robust technical foundation.

The platform supports multiple programming languages including Python, JavaScript, and PHP through RESTful API endpoints. Authentication requires a JWT token issued by Nyckel's identity provider, enabling secure integration into existing workflows.

Key Features

Nyckel offers several pre-trained classifiers covering a range of image and text analysis tasks, in addition to tools for building custom classifiers with minimal expertise. Their existing models include gender identification from hand images, product color recognition, and logo detection, while their platform allows users to create custom classifiers quickly through automatic testing and active learning capabilities.

The company's technical infrastructure enables rapid model development without requiring PhD-level expertise. Developers can create custom classifiers using just minutes of input data, with automatic testing capabilities evaluating hundreds of potential models against user datasets. This approach streamlines the development process while maintaining robust accuracy standards.

Pre-trained Models

Nyckel's pre-trained models address specific use cases through carefully curated datasets. The company offers a gender identification tool trained on hand images from the Hands and Palms Dataset, identifying male and female hands with confidence scores indicating prediction reliability. For logo detection, the platform utilizes a specific dataset containing brand logos, allowing users to create custom brand identifier models with minimal input.

Custom Classifier Development

Building upon their technical foundation, Nyckel enables developers to create custom classifiers through several key features. The platform automatically tests hundreds of potential models against user data, providing an efficient way to develop reliable classification systems. Active learning capabilities further enhance model accuracy by identifying specific areas for improvement, allowing for iterative refinement of custom classifiers.

API and Integration

Nyckel provides a comprehensive set of RESTful HTTPS endpoints for working with classification functions. The API supports multiple programming languages including Python, JavaScript, and PHP, with authentication requiring a JWT token issued by Nyckel's identity provider.

Function management operations include retrieving aggregate information about functions through GET GET /v1/functions/<functionId>/summary, which provides details on total samples and annotated samples broken down by label ID. Developers can perform CRUD operations on samples using operations like DELETE DELETE /v1/functions/<functionId>/samples/<sampleId>/annotation to manage specific sample annotations.

Function invocation follows a straightforward pattern with POST /v0.9/functions/<functionId>/invoke, which takes input text and returns an array of Prediction objects representing associated labels. Each Prediction object includes labelId, labelName, and confidence values indicating the model's certainty about each prediction.

The system supports active learning by automatically identifying areas where model accuracy can be improved, allowing for iterative refinement of custom classifiers. Nyckel employs robust security measures including rate-limiting at 25 requests per second and concurrency limits of 25 requests, with these limits relaxed for Enterprise plans.

All API requests require an Authorization header containing the JWT token, with supported methods including curl, Python's requests module, JavaScript's fetch method, and PHP's cURL functions. Response structures vary based on the specific endpoint, with function summaries including counts of total and annotated samples, while individual predictions contain labelId, labelName, and confidence scores.

Security and Compliance

Nyckel employs several security measures to protect user data. All requests to their API must include an Authorization header containing a JWT token issued by their identity provider, with supported methods including curl, Python's requests module, JavaScript's fetch method, and PHP's cURL functions. The system implements rate-limiting at 25 requests per second and concurrency limits of 25 requests, with these limits relaxed for Enterprise plans. Exceeding these limits results in a 429 Throttled error code.

The company maintains strict data handling practices that prevent internal data sharing. They use imported and annotated data to train their functions, with guidelines recommending providing data similar to what the function will encounter in production, maintaining balanced data across classes, and increasing data volume when training functions. Cross-validation methodology involves training multiple models on most but not all samples, with each model predicting labels for the remaining samples to provide a comprehensive accuracy assessment.

Function accuracy is displayed in the left navigation panel of the console, with the top bar reflecting overall accuracy (number of correctly predicted samples divided by total samples) and class level accuracy bars showing performance across different categories. The system handles both URL and data URI inputs for image classification, returning predictions in an array of Prediction objects containing labelId, labelName, and confidence scores.