2026AI ecommerce intelligence platform
Fidras
A Shopify-focused SaaS that folds the whole growth stack into one brain: customer analytics, CLV and RFM segmentation, churn prediction, sales forecasting, product recommendations, abandoned-cart recovery, an AI chatbot, multi-channel marketing and ad intelligence.
Status · Live product — fidras.com. Shopify-focused platform built at CatalystIQ.
The problem
A Shopify merchant doing real volume has analytics in one tool, email in another, ads in a third and churn nowhere. Each tool reports activity; none of them answers the operator's actual question, which is what to do next week and to whom. Smaller merchants cannot afford the data team that would join it all up.
The approach
- 01Unify order, customer and event data from the store into one model so every downstream feature reads the same truth.
- 02Compute the quantitative backbone: RFM segmentation, CLV, churn probability and sales forecasts.
- 03Translate model output into actions — a segment to campaign, a cart to recover, a product to recommend, a budget to shift.
- 04Deliver the actions where they execute: email, WhatsApp, SMS, on-site recommendation and an AI chatbot, with pixel and event tracking closing the measurement loop.
What I did
- Designed the multi-tenant data model and the Shopify ingestion layer.
- Built the modelling stack: RFM and segmentation, CLV, churn prediction and sales forecasting.
- Built the recommendation and abandoned-cart recovery flows and the campaign execution paths.
- Built the application surface and wired ad and pixel intelligence into the same reporting model.
Architecture
Tap any component to see what it does and what it runs on.
Ingestion
Processing
Models
Orchestration
Storage
Interfaces
Select any component to inspect it. Architecture as implemented.
Trade-offs I would defend
Interpretable models over the maximum score
Gradient boosting and RFM beat a deep model here, not on benchmark accuracy but because a merchant will only act on a prediction they can interrogate. The segment has to be explainable in one sentence.
One data model, many features
Every feature was pushed through a single canonical schema rather than shipping faster with per-feature queries. Slower at the start; the only reason churn, ads and recommendations agree with each other now.
Stack & evidence
- Shopify API
- Python
- scikit-learn
- LightGBM
- Postgres
- Next.js
- Live site
- fidras.com
- Role
- Founder / engineer — architecture and modelling
- Domain
- Ecommerce / Shopify
Next case study
Voice Agent "Mariam"
Enterprise conversational AI over the phone