Architecture

How Geminus works

The full pipeline, from brand objective to causal lift measurement. Click any stage for details.

Objective Brand objective Plain-language growth goal AI Translation AI strategy translation Agentic LLM layer: NTB, lapsed, frequency, conquest Audience Definition Audience definition Labels, eligibility, outcome window Ground Truth Seed & label extraction Positive & negative labels from the panel, per audience Data Sources Data sources Retail transaction panel Credit card transaction panel DTC purchase data Household Signal Household history as text Full purchase history, plain language. Not aggregated, not compressed. Feature Prep Item & category resolution LLM: product, brand, category. Cached, reused across every strategy. Model Training Fine-tuned compact model Open-weight, self-hosted, LoRA fine-tuned per strategy No real-time constraint: no distillation, no teacher/student split Scoring Batch scoring, full household universe Bin prediction, converted to score via empirical rate table Audience Output Audience output Scored, ranked, and sized Holdout Holdout split Random split before activation Activation Activation Platforms of client's choice Measurement Causal lift measurement Treatment vs. holdout, unmodeled

Click any stage for details

If the objective matters, the starting point matters.

Talk to us about starting your next campaign from the objective.

Start a conversation

Or write directly: