Case study

Winglet

A personal-shopping storefront: find a product in a European online store, see the full landed cost, order in a couple of taps. Built on Cloudflare’s free tier — one Worker, no database, no build step.

Cloudflare WorkersServerless, no DBLLM classificationTelegram botCron automationVanilla JS
The problem

Many European stores don’t ship abroad, and resellers quote blind — “we’ll tell you the price later,” with no idea what delivery of a specific item will cost.

Winglet estimates the full landed cost before the order — per found product, against the carrier’s real tariff.

What the platform does
Product search

Search over Google Shopping, results normalised to EUR. Russian queries work: a Cyrillic query is run through an LLM that translates it and canonicalises brand names (“стон айленд” → “Stone Island”) before the search.

Landed cost on every card

Item price plus a live delivery estimate from the carrier’s tariff — per-kg categories, per-item rates, a minimum charge, volumetric weight, a high-value surcharge — and a configurable platform commission, folded into a single “with delivery” figure.

LLM product classification

A model resolves type → tariff category, condition, packaging, weight subtype; that drives the shipping math. If the model is slow or down, it falls back to keyword heuristics and a manual category pick. It never blocks the request.

Product detail page

Classification plus a second search that finds the direct merchant link — domain-matched, distinguishing “open in store” from “similar on sale”. Editable cost breakdown, size presets and a free-text field.

Cart & batch quote

A multi-item cart in browser storage. The batch quote applies the carrier’s shipment-level rules — minimum charge once per shipment, surcharges across the whole list — and collapses identical item + size into quantities.

Order → Telegram

Checkout posts the order — items, sizes, contact, a draft breakdown for the manager — straight to a Telegram channel via a bot. No database anywhere; orders live in Telegram.

Client vs operator modes

A ?admin=1 flag reveals operator tooling — categories, weights, markup %, a paste-a-URL parsing tab, the full breakdown. The default client view hides all of it and filters operator-facing warnings.

Worth a closer look
Nice touches
Discount detection with no discount data

Google Shopping doesn’t expose original or strikethrough prices. Winglet infers value two ways: price against the median of the whole result set, and an independent LLM estimate of the model’s normal retail price. A guard stops a wrong “used” label from killing a real sale.

A deals feed that markets itself

A “Deals” tab rotates a pool of “brand + sale” queries, keeps items well below median that aren’t used or suspiciously cheap, computes landed cost, dedupes and sorts by discount. Cached, rebuilt daily by a Cron Trigger that also auto-posts a digest of the top finds to a Telegram sales channel — a zero-cost acquisition loop.

Currency toggle € / ₽ with no refetch

The rate is fetched once — Russian Central Bank, with a fallback source — and cached. Toggling re-renders every price on the page client-side, with zero new requests.

Paste-a-URL parsing (operator)

The Worker fetches the page and reads JSON-LD / Open Graph; Shopify and WooCommerce parse cleanly. It detects bot-blocked or JS-only pages and returns a clean “enter manually” instead of letting the model hallucinate.

A timeout on every external call

Search, LLM and FX calls are wrapped in an AbortController — a hung free-tier model can’t hang a request.

Architecture
One Worker, free tier, no database

One Cloudflare Worker on the free tier plus a static vanilla-JS page. No framework, no build step, no database — the system is stateless: everything is computed on request or lives in Telegram.

  • Cache API for the FX rate and the deals feed.
  • Cron Triggers for the daily deals rebuild and auto-post.
  • Secrets for the external API keys.
  • Fits entirely inside the free tiers — free Worker, free LLM model, no paid search plan.
  • Works on the default workers.dev domain, no custom domain needed.
What I took from it

Building a whole product on free infrastructure forces the call on where complexity is actually worth it and where the boring thing will do: no database, cache instead of storage, cron instead of a queue, graceful degradation instead of an error when a free dependency falls over.

Building on constraints, or from scratch?