Model card
Which models power this platform, and why they are open.
Models in use
| Model | Role | Weights |
|---|---|---|
| gpt-oss-120b (OpenAI) | Primary generation: segmentation, campaigns, conversational strategist, post formatting | Open weights (Apache 2.0) |
| Gemini 2.5 Flash / Pro | Alternate route, selected per request for capacity and fail-over | Proprietary (Google) |
gpt-oss-120b (OpenAI)
Primary generation: segmentation, campaigns, conversational strategist, post formatting
Open weights (Apache 2.0)
Gemini 2.5 Flash / Pro
Alternate route, selected per request for capacity and fail-over
Proprietary (Google)
Why open-weight models
- •Open-weight models provide an additional option for controlling inference costs and supporting a free core service.
- •A provider abstraction reduces dependence on a single hosted model, although providers can still change prices and availability.
- •The model, its weights and its behavior can be audited, which the platform's explainability commitments depend on.
How models are selected
A provider-abstraction layer means a model can be substituted by configuration rather than a rewrite.
The codebase includes a model-comparison harness. A complete marketing-quality evaluation across all candidate model families has not yet been established.
Explainability
Prompts request reasons for recommendations; this alone does not validate their truth. For recognized CSV date/order-value fields, code computes and preserves cohort counts and rules independently of generated marketing suggestions. Other inputs and all creative suggestions still require review.
Known limitations
- •Outputs are generated text and should be reviewed before publication.
- •Benchmark segments for a business type and location are estimates, not measurements of that specific business.
- •The platform does not have access to a business's live ad performance unless the owner provides it.
- •Generated figures and statistics in marketing copy should be verified by the owner.
Training data
We do not run a training or fine-tuning pipeline on uploads. Analysis is sent to configured third-party inference services; their handling is governed by applicable service terms and account settings.