admesh/agentic-intent-classifier

Ingence 0.1

Ingence is a proprietary intent-understanding model developed by AdMesh for conversational and agentic applications.

Release status: Experimental. Ingence 0.1 is intended for evaluation and controlled use. It has not been validated for safety-critical or high-impact decisions.

What It Does

Ingence converts a user request or short conversation into structured signals describing:

• —the user's general intent;
• —a more specific intent subtype;
• —the user's stage in a decision journey;
• —a relevant contextual content category; and
• —whether the system should fall back because a prediction is uncertain.

These signals can support intent routing, contextual relevance, aggregate analytics, and controlled advertising or content-selection experiences.

Intended Uses

Ingence 0.1 is intended for:

• —intent routing in assistants and agents;
• —understanding research, comparison, and purchase-oriented requests;
• —contextual content and advertising selection;
• —aggregate intent analytics; and
• —research, testing, and controlled product evaluation.

Restricted and Unsupported Uses

Ingence must not be used:

• —to infer sensitive personal characteristics;
• —for surveillance or individual behavioral profiling;
• —as the sole moderation or safety mechanism;
• —to make decisions about credit, employment, housing, insurance, healthcare, education, or legal services;
• —as the sole basis for decisions that materially affect a person; or
• —to target advertising using sensitive personal information.

Training and Evaluation

Ingence 0.1 was developed using a mixture of synthetic, curated, and taxonomy-based examples. Synthetic examples were used to improve coverage of different intents and linguistic patterns.

Important qualifications:

• —Synthetic-label agreement does not establish real-world accuracy.
• —Training data may not represent every population, dialect, industry, or interaction style.
• —Evaluation on representative, human-reviewed production data is ongoing.
• —Definitive accuracy, fairness, and calibration claims are not currently published for this experimental release.

Confidence Scores

Model confidence values are not guaranteed probabilities. Applications should not interpret a score such as 0.80 as an 80% probability that a prediction is correct. Integrations should honor uncertainty and fallback indicators.

Known Limitations

Ingence 0.1 may:

• —confuse closely related research and decision stages;
• —misclassify short or context-dependent follow-up messages;
• —have difficulty distinguishing general price questions from commercial intent;
• —return broad content categories when several related categories apply;
• —fall back on valid requests when confidence is low;
• —perform differently on language styles not represented in development data; and
• —produce unreliable results for languages other than English.

Outputs are probabilistic predictions and should be treated as one input into a broader decision system.

Safety and Privacy

Applications using Ingence should:

• —minimize personal information included in requests;
• —avoid retaining raw conversations unnecessarily;
• —provide a safe fallback when the service is uncertain or unavailable;
• —use independent policy enforcement for sensitive applications;
• —monitor model quality and distribution drift; and
• —clearly disclose when automated predictions materially influence an experience.

Access and Intellectual Property

Ingence 0.1 is proprietary technology owned by AdMesh. Model weights, source code, training data, internal evaluation artifacts, and deployment details are not publicly distributed. No license to copy, modify, redistribute, reverse engineer, or create derivative works is granted by this model card.

Authorized access may be provided through AdMesh-operated services under applicable commercial terms.

Version Information

Contact

For product questions, feedback, privacy requests, or responsible disclosure, visit useadmesh.com.