Annie uses a hierarchical mixture-of-experts (MoE) architecture. Within this architecture, a pipeline role is a defined function in the processing chain — a specialist slot filled by whichever model is best suited to that task.
Annie ships with twelve open pipeline roles today, and the architecture is designed to add more as deployments grow — it is not a fixed ceiling. Each role has a defined interface (inputs, outputs, confidence thresholds), but the model filling it can be swapped, replaced, or fine-tuned without touching the rest of the platform. Roles include functions like initial parsing, domain-specific reasoning, compliance checking, summarisation, and output verification.
This architecture means Annie is not dependent on any single model or provider. You can bring your own models for specific roles, deploy open-weight Workforce Foundation Models, or use fine-tuned variants trained on your data — all within the same pipeline.
Workforce Foundation Models (WFMs) are Annie's suite of sovereign pre-trained models — trained from scratch on curated, licensed data with no external model weights inherited from any frontier AI provider.
Key properties of the WFMs:
- Sovereign by origin — trained and hosted on infrastructure controlled by Evari and its partners; no proprietary weights from OpenAI, Anthropic, Google, or Meta
- Regulated-industry focus — pre-trained on corpora relevant to government, financial services, healthcare, and legal sectors
- Deployable on-premise — model weights are provided to customers and run locally; no API calls required
- Fine-tunable — WFMs serve as the starting point for Cognition Stream fine-tuning on your task data
You can deploy Annie using WFMs alone, bring your own models into pipeline roles, or blend both approaches.
The Judgment Panel is Annie's multi-model consensus and verification layer. For significant outputs — decisions, recommendations, summaries with material consequences — Annie does not rely on a single model's output. Instead, the Judgment Panel routes the task to multiple specialist models, collects their independent outputs, and applies a structured consensus algorithm.
The result includes:
- The accepted output
- Which models were consulted and their individual responses
- Whether consensus was reached or a threshold was applied
- Confidence scores and dissenting signals
This structured record forms a verifiable audit trail — a critical requirement for regulated industries where decisions must be explainable to regulators, auditors, or courts.
The Judgment Panel cannot be bypassed for outputs above a configurable confidence threshold. This is a platform-level control, not an optional feature.
The Cognition Stream is Annie's continuous learning mechanism. During off-peak periods — typically overnight — Annie analyses accumulated task data, identifies patterns where its outputs were corrected or rated poorly, and runs a targeted fine-tuning cycle to improve performance on those task types.
Key properties:
- Runs on your infrastructure — fine-tuning happens locally; training data never leaves your environment
- Low cost — a Cognition Stream run costs less than $5 in compute
- Targeted improvement — fine-tuning is scoped to specific pipeline roles and task types, not wholesale model retraining
- Auditable — each Cognition Stream run produces a log of what was updated and why; you control fine-tuning schedules
Over time, the Cognition Stream means Annie gets progressively better at your specific workflows — policy review, contract analysis, claims processing — without relying on generic updates from a third-party provider.
Yes. Annie's pipeline architecture is open — each of the pipeline roles (twelve today) accepts any model that conforms to the role's interface specification (input format, output schema, confidence scoring protocol).
You can bring:
- Open-weight models — Llama, Mistral, Falcon, Qwen, and other models with permissive licences
- Fine-tuned variants — models you have trained or fine-tuned on your own data
- Domain-specialist models — sector-specific models such as clinical NLP or legal document analysis
- Internal models — proprietary models your organisation has developed
Models do not need to be publicly available or sourced from any particular provider. They need to meet the role's interface specification, which Annie validates at deployment.
You can mix Workforce Foundation Models and your own models across different pipeline roles — for example, using a WFM for general reasoning and a domain-specific model for compliance checking.