How we come up with the right price.
FreshScale AI is a machine-learning ensemble. We combine large language models from every major frontier lab with classical forecasting to turn expiring produce into the exact markdown that clears the shelf — without giving away margin.
Primary reasoning engine — evaluates each SKU's expiry, cost floor, and demand signal to propose the markdown.
Backup reasoner and structured-output cross-check when Gemini's confidence dips below threshold.
Long-context validator — reads a store's 90-day markdown history to spot patterns and outliers.
Live retrieval — pulls local weather, event, and competitor-promo signals to condition the forecast.
Open-weight fallback used for on-prem pilots where data can't leave the retailer's network.
Cost-efficient batch scorer that pre-ranks tomorrow's expiring SKUs overnight.
Product names and logos are trademarks of their respective owners. Shown for illustrative purposes to describe the model families FreshScale AI's ensemble draws on.
Every markdown is the output of a five-step pipeline running in your store, in real time.
Point-of-sale, inventory, weather, day-of-week, local events, and competitor promotions stream into the model.
A gradient-boosted demand model predicts sell-through by hour for every expiring SKU.
The LLM ensemble weighs the forecast against cost floor, brand rules, and confidence to draft a markdown.
Guardrails enforce the price never drops below cost and derive the discount % from the actual price change.
Real outcomes — units sold, revenue, waste avoided — flow back into training so tomorrow's price is smarter.
No single model gets fresh-produce pricing right on its own.
Gemini is fast and cheap. GPT is a strong structured reasoner. Claude reads long histories without losing the thread. Perplexity keeps us grounded in what's happening outside the store right now. Blending them — plus a purpose-built demand forecast — is how we hit 100% sell-through without ever surging a price on a shopper.
See it on your data →