Artificial Intelligence (AI) Questions

Which products are enabled by Claravine AI?

Claravine AI technology enables several different features and capabilities. 

Which categories of AI are used in Claravine AI?

Claravine AI technology solves different needs and workflow complexity using different features and capabilities. Two categories of artificial intelligence are available for use in Claravine products, Computer Vision (“CV”) and Natural Language Processing (NLP).

Computer Vision is used for:

• Content Comprehension: AI scans and analyzes visual content to identify, classify, and label objects and characteristics in images and videos.
• Content ID: AI analyzes an image or video, constructs the aboutness of that asset, then generates a Content ID that becomes uniquely tied to that asset and similar assets. 

Natural Language Processing, more specifically with large language models (“LLMs”), is used for:

• Taxonomy Translation: AI automates the transformation and alignment of different taxonomies across regions, business units or clients.
• Smart Correct: AI automatically suggests fixes across errored metadata, making corrections as simple as a click and quick review. 
• Smart Campaign Builder: With a prompt containing your campaign objectives, the Smart Campaign Builder generates thousands of data standard-compliant campaign names and IDs in seconds, based on customer-defined taxonomy standards.

How do we train models?

Currently, we use a combination of synthetic data, public data and examples modeled after real-world cases.

• The training and testing datasets:
• Use generic & synthetic data examples
• Use open sourced libraries and datasets
• Do not use customer data
• Do not use PII

Which LLM is used in NLP capabilities?

Claravine uses LLMs via 3rd party API calls to secure foundation models available on the market: 

Customer data is not used to train or make improvements to these models. 

Claravine uses an LLM proxy for the platform to route to a pre-configured LLM based on:

• specific feature or account need
• adapting to current and future LLMs, as well as contractual requirements
• campaign size that may require different context window

How does Computer Vision work?

When using our Computer Vision technology, Claravine is able to scan creative content such as images and video assets. Using our patented CV, we deliver content comprehension and content similarity identifiers using frame-by-frame Artificial Intelligence analysis.

When videos are processed, classification and confidence analyses are applied at the scene level as well as aggregated across scenes into a global classification summary for the video. Unlike metadata scraping, Claravine’s approach does not rely on self-declared and often inconsistent labels that typically prevent or mislead deep contextual analysis.

When determining similarity, we leverage an Approximate Nearest Neighbor algorithm to facilitate the association of content taxonomy analyzed results to the content in question.

What data is processed by AI features?

• Taxonomy Translation: Claravine ingests two types of input data - media trafficking sheets / data dictionaries and list-value sheets. These datasets are provided by an end-user in Claravine. Both data input and output remain accessible in the customer account only.
• Smart Campaign Builder: Claravine ingests a prompt provided by an end-user in Claravine. The AI feature then ingests the associated taxonomy configuration in the customer account. Both data input and output remain accessible only in the customer account.
• Smart Correct: Claravine ingests a campaign dataset selected by an end-user in Claravine. The AI feature then ingests the associated taxonomy configuration in the customer account. Both data input and output remain accessible only in the customer account.
• Content Comprehension and Content ID: Claravine processes images and videos associated with dataset rows that are submitted by the customer from a customer-configured integration securely connected to the Claravine platform. Both data input and output remain accessible only in the customer account.

How is data safeguarded?

Claravine uses a multi-tenant architecture that enforces strict data separation by account ID, ensuring that each organization’s data is isolated and private. 

How does Claravine measure and maintain accuracy over time?

We measure accuracy using a combination of precision and recall, with an emphasis on user agreement. For example, translations are scored based on how many of the model’s suggestions are accepted by users within a customer account.

To maintain accuracy over time, we regularly evaluate performance at both the organization level and the global level. This helps us detect any model drift or degradation in performance.

Claravine remains proactive by benchmarking accuracy against newly released LLMs and LLM versions, ensuring we are leveraging the most effective and up-to-date models available towards improvements that are clear and measurable.

What other relevant resources can I reference?

Claravine Privacy Policy