AI Transparency Statement

Last Updated: January 2025

Preface

Transparency is essential to building and maintaining trust in artificial intelligence systems. This AI Transparency Statement explains how Synetecs, Inc. ("Synetecs," "the Company," "we," "us," or "our") designs, trains, and operates its AI systems, as well as their known capabilities and limitations.

This statement applies to all AI-powered products, platforms, models, and services provided by Synetecs.

1. Model Architecture

Synetecs employs a hybrid neuro-symbolic architecture that combines statistical machine learning with deterministic logic systems.

• Large Language Models (LLMs): Natural language understanding and generation
• Symbolic Reasoning: Rule-based logic, constraints, and policy enforcement
• Specialized Models: Speech recognition (ASR), text-to-speech (TTS), and optical character recognition (OCR)
• Context Store: Vector-based semantic memory for retrieval and long-term context

2. Training Data

2.1 Data Sources

Our models are trained using a mixture of the following data sources:

• Publicly available datasets
• Commercially licensed datasets obtained with appropriate rights
• Synthetic data generated for specific tasks and edge cases
• Customer-provided data only where explicit opt-in consent has been granted

Customer data is never used for model training or improvement without explicit authorization.

2.2 Low-Resource Languages

For low-resource languages, including Dhivehi, Synetecs works with native speakers, cultural institutions, and partner organizations to collect high-quality data.

All data collection follows ethical guidelines, respects cultural context, and ensures fair compensation for contributors.

3. Model Capabilities and Limitations

3.1 Capabilities

Depending on configuration and deployment, Synetecs AI systems may support:

• Natural language understanding and generation across multiple languages
• Context-aware, multi-turn conversational workflows
• Task planning, orchestration, and structured reasoning
• Document analysis and information extraction
• Speech recognition and speech synthesis

3.2 Known Limitations

AI systems are probabilistic and subject to limitations, including:

• Hallucinations: Outputs may appear plausible but be factually incorrect
• Bias: Residual bias may exist despite mitigation efforts
• Context Length: Performance may degrade with very long inputs or sessions
• Domain Sensitivity: Performance is strongest within trained domains
• Temporal Knowledge: Models do not have real-time awareness unless explicitly integrated

4. Human Oversight

Synetecs recommends and, in some cases, requires human oversight for AI-assisted use cases, including:

• Legal, medical, financial, or safety-critical decisions
• Processing of sensitive or personal data
• Novel or out-of-distribution scenarios
• Applications subject to regulatory or audit requirements

5. Bias and Fairness

5.1 Bias Testing

Models are evaluated for bias across demographic attributes such as gender, age, and other relevant characteristics.

When bias is detected, Synetecs applies both technical mitigations (such as data balancing or debiasing techniques) and procedural controls (such as human review).

5.2 Ongoing Monitoring

Deployed models are continuously monitored for performance and fairness degradation. Periodic audits assess subgroup outcomes, and models are updated or retrained when material disparities are identified.

6. Model Updates and Versioning

Synetecs regularly updates its AI models to improve performance, address identified issues, and incorporate safety improvements.

Major updates are communicated in advance where feasible. Deprecated model versions remain available for a limited period, typically ninety (90) days, to support customer transition.

7. Explainability

For high-impact or regulated use cases, Synetecs may provide additional explainability mechanisms, including:

• Confidence or uncertainty indicators
• Input feature attribution where technically feasible
• Decision traces for multi-step reasoning workflows
• Counterfactual analysis illustrating alternative outcomes

8. Environmental Impact

AI model training and inference consume computational resources and energy. Synetecs optimizes models for efficiency, prioritizes responsible infrastructure use, and leverages architectural approaches that reduce unnecessary computation.

Where possible, renewable energy sources and carbon offset programs are used to mitigate environmental impact.

9. Third-Party Audits

Synetecs engages independent third parties to assess AI systems for bias, safety, and compliance with applicable standards.

Audit summaries or reports may be made available to enterprise customers under appropriate confidentiality agreements.

10. Contact Information

Questions regarding Synetecs AI systems, transparency practices, or responsible use may be directed to: