NaviTo — AI Classifieds for Expats

# openrouter_analyzer.py — primary model and fallback chain
DEFAULT_MODEL = os.getenv("OPENROUTER_MODEL", "google/gemma-4-31b-it:free")
FALLBACK_MODELS = ["google/gemma-4-26b-it:free", "google/gemini-2.5-flash-lite"]
# if every model fails — a Telegram alert to the admin listing the models tried

04: Catalogue and Listing

Once the chaos is turned into PostgreSQL rows, the React front end takes over: a catalogue with category, city and price filters, a listing page with a gallery, seller contact and a link to the original Telegram message. Photos live in Cloudflare R2 rather than the database — that removed DB transfer and sped up pages.

Three markets with their own currency and pricing logic — Thailand (฿), Georgia (₾), Vietnam (thousand ₫); the region is detected by IP on first visit.

export type MarketSlug = 'thailand' | 'georgia' | 'vietnam';
const CURRENCY_SYMBOLS = { THB: '฿', GEL: '₾', K_VND: 'тыс. ₫', USD: '$', RUB: '₽' };

05: For Sellers — an AI Assistant Instead of a Form

Sign-in is a Telegram deep link, no phone number. A listing doesn't have to be filled in — it can be told: the user drops photos, the AI recognises the item, asks clarifying questions with quick-reply chips (condition, what's included, area), suggests a recommended price range for the market — and assembles the card. The chat uses a separate model from the parser.

Subscriptions and notifications

A buyer subscribes to a category in their city and gets new listings in the Telegram bot and via web push (PWA, VAPID). VIP listings rise to the home page; national VIP shows in every city of the country.

06: Admin and Data Hygiene

• Chat sources with per-chat parsing pause
• Keyword blacklist and author blacklist (their listings auto-archive)
• AI log: processing history per message with model, response and errors
• Premium users, VIP, seller reassignment, batch re-categorisation of "misc" (within Vercel's 60-second limit)
• Auto-purge: archived listings are deleted after 10 days, with a CSV export to Telegram first

07: Stack

• Front: React 19, Vite, TypeScript, Tailwind, PWA (service worker, web push), React Router
• Back: FastAPI on Vercel Serverless, SQLAlchemy, PostgreSQL, Cloudflare R2
• Listener: Python + Telethon on Railway (long-running process)
• AI: OpenRouter (Gemma 4, Gemini Flash Lite); locally Ollama + Qwen 3 14B for development

08: Promo and Outcome

The /ad landing — "Avito for our people abroad" — explains the product in one screen.

The architecture removes the flea-market moderator as a profession: the catalogue builds itself, sellers fill in nothing, buyers stop scrolling chats. The platform scales to a new country by adding a market and a list of chats.