One Shop, Its Owner and an Agent. One Predictive Database.
PetNord, a Nordic pet store where every screen runs on Aito's predictions, with no model training and no retraining schedule
The shop a shopper sees, the owner's decisions (what to put on offer, at what price, what to reorder, who is about to leave) and an agent that reads the same store memory, all on one Aito database. Every view shows the query behind it. PetNord's data is generated, with planted effects. It shows how Aito works, not how much it would lift a real shop; the real-data result is in the Evidence section.
The Site, by Who Asks
The shopper's Shop, the owner's views, two back-office tools, and the evidence that says how well. Pick a view to read how it is built on GitHub.
Shop
What a shopper sees. A shop knows what a visit is for, so it shows one view per intent, from what the shopper told it, the month and the cart.
Run the Shop
The owner's views. Each leads with the decision it supports, its estimated impact and the reasons from $why, and links to the evidence that says how well.
Product
The funnel from showings to clicks, purchases and profit, beside the month as it was counted. Change the offer, list price, stars, pictures, description or advertising and the funnel moves
See in demo →Source →
Pricing
An editable list price, the demand range and the profit at each price, only within the prices the product has sold at
See in demo →Source →
Retention
Who is about to leave and why, ranked by the spend at stake. For each customer the shop's actions (an email, a discount, a voucher, an ad, an apology) are ranked against doing nothing, after what each costs, with the signs that came before churn
See in demo →Source →
Supply
What to reorder now and what is stuck on the shelf, from next month's forecast per product, with the euros that follow from it
See in demo →Source →
Patterns
What goes in the same basket, what goes with a purchase and who buys what, each lift with its interval
See in demo →Operate
Back-office tools. Several fields predicted in parallel, each with its $why.
Evidence
How well. Accuracy numbers live only here, each page marked Real or Synthetic data.
Business Reorders (Real data)
On UCI Online Retail II, for buyers with 1–2 earlier orders, Aito's top suggestion was in their next orders 31 % of the time, against 15 % for the strongest of five classical baselines
See in demo →
Model Checks (Synthetic)
Held-out _evaluate runs on PetNord, including the ones that fail
See in demo →Source →Implementation Details
Explore the source, the demo moments, and the architecture decisions
{
"from": "impressions",
"where": { "product_sku": "SKU-PT-0563", "product_state.discount_band": "20+" },
"predict": "purchased",
"select": ["$p", "feature", "$why"]
}
Why This Pattern Works for E-commerce SaaS
One predictive database for the shop, the owner and the agent, every change learned from as soon as it is written
SQL-Like Predictions
- _predict, _relate, _recommend, _estimate, _evaluate, _search
- No feature engineering required
- Confidence + $why on every prediction
Consumer + Operational, One Substrate
- Shop, search, kit suggestions, buy again
- Product funnel, pricing, supply
- Retention and win-back, review triage
Production-Ready Architecture
- Next.js + FastAPI, one service per view
- Only what a shop observes, read through links with no joins
- Levers write to a throwaway sandbox branch, never the live store
Address
Episto Oy
Putouskuja 6 a 2
01600 Vantaa
Finland
VAT ID FI34337429
