sellerbox.ai

A production SaaS for Wildberries marketplace sellers: a Telegram bot and Mini App. The core is an AI card-audit engine: for a given article it pulls WB-API data and returns a 0–100 score across ads, infographics, SEO, reviews and price, with infographic and description generation, review replies and a unit-economics calculator alongside it. This is a full-stack product with billing, not a single-model demo: bot, Mini App, admin panel and credit-based payment.

  • An article → a 0–100 report: ads, infographics, SEO, reviews, price.
  • Fixes sit next to the diagnosis: infographics, descriptions, replies.
  • A real SaaS with credit-based billing, not a prototype.
Problem

A seller cannot see what to fix on a card first.

A seller has dozens of cards and no clear signal about what actually drags each one down: ads, weak infographics, thin SEO, bad reviews or price. The data is scattered across the WB API and third-party tools; diagnosing one card by hand is slow, and doing it across a whole catalog is impossible.

  • Dozens of cards — and no clear signal about what to fix first.
  • The data is scattered across the WB API and various tools.
  • Diagnosing one card by hand is slow; a whole catalog is unrealistic.
  • Ads, infographics, SEO, reviews and price all pull at once.
Solution

A 0–100 score per card — and the tools to raise it.

For a given article the product pulls WB-API data into one report scored 0–100 across five axes, and attaches the fixes right next to the diagnosis: infographic and description generation, review replies and a unit-economics calculator.

The audit report: five axes — ads, infographics, SEO, reviews, price — and one overall score.
Top queries with positions: you can see at once where the card sinks in search.
A generated card: the base render and the text live on separate layers, up to 4K.
Another generated infographic — each niche gets its own style.
Unit economics in the Mini-App: WB commissions, logistics, taxes and marketing — down to per-unit profit.
  1. 01

    Fold a card into one report

    For a given article we pull WB-API data and score it 0–100 across five axes: ads, infographics, SEO, reviews, price. Two formats: a short card right in the bot and a full report with charts.

  2. 02

    Score the ads

    CTR, CPC and CPM, keyword clustering, and minus-word recommendations.

  3. 03

    Score infographics with a vision model

    Every card photo gets a score — what exactly to improve in the shot.

  4. 04

    Break down reviews and price

    Rating, buyout rate, the cart → order → buyout funnel and complaint mining; price versus competitors and 30-day revenue.

  5. 05

    Provide the fixes

    A card is generated in three steps: product → description → prompt and render, with the text laid on top as a separate layer. Plus SEO-oriented AI descriptions, review replies and a unit-economics calculator.

Technical

A full-stack SaaS with an AI + vision pipeline and billing.

A Python/FastAPI backend and an aiogram bot, Postgres/Neon + pgvector for RAG, SQLAlchemy/Alembic, a Redis + ARQ task queue. The client is a Telegram Mini App with server-verified initData; the admin is server-rendered on FastAPI + Jinja2 (deliberately no React-admin — easier in AI-assisted development). Payment is credit- and subscription-based.

The AI layer combines OpenAI, Claude and Gemini; a dedicated vision model scores card photos, 4K infographic generation runs through Nano Banana Pro and GPT Image 2, and video through Kling. The unit-economics engine matches industry-standard calculators (FBO/FBS commissions, logistics, SPP, KZ tax regimes, live per-category tariffs). Infra is AWS (EC2/S3/CloudFront/Route53), Cloudflare, Docker Compose with GitHub Actions autodeploy, and Sentry.

01

FastAPI backend and an aiogram bot.

02

Postgres/Neon + pgvector (RAG), a Redis + ARQ queue.

03

Telegram Mini App with server-verified initData.

04

A vision model — per-photo scoring of the card.

05

4K infographic generation (Nano Banana Pro / GPT Image 2), video via Kling.

06

A unit-economics engine matching industry-standard calculators (FBO/FBS, SPP, KZ taxes, tariffs).

07

Credit and subscription billing, GitHub Actions autodeploy, Sentry.

08

A single audit costs about a kopeck to run; abuse is capped at one audit per card per hour.

Stack
Python / FastAPI backend and audit engine
Telegram Mini App the seller client
Neon Postgres + pgvector data and RAG
Redis + ARQ background jobs
Vision model per-photo card scoring
Nano Banana Pro · GPT Image 2 4K infographics
Unit economics FBO/FBS, SPP, tariffs
AWS infra and deploys
Outcome

What changed

01

Card audit became one button

For a given article — a 0–100 report across five axes instead of a manual sweep through five tabs.

02

Diagnosis and fixes sit together

Vision photo scoring, 4K infographic generation and AI descriptions come right after the report, not as a separate tool.

03

A production SaaS with billing

A full-stack product — bot, Mini App and admin — with credit-based payment, not a prototype.

Need a full-stack AI product with auditing, generation and billing?

We build these SaaS end to end: API data collection, an analytics engine, an AI and vision pipeline, a Mini App, and credit-based payment.

Brief (optional)