All projects

dunlimited

Singapore's top-ranked marketing agency and its family of owned product brands.

Work

AI Support for Shopify Stores

dunlimited03/2026 – 04/2026
660/mo
Tickets handled
5
Shops

The AI drafts every reply and sends it directly when confident enough; otherwise a human just approves it instead of writing from scratch. It filters spam, fills templates from Shopify order context, and escalates uncertain cases to a human via Telegram, backed by guardrails tested during development.

One store, everleakproof (lingerie), is non-returnable once shipped, so the bot's instant reaction time was the only thing that could still catch a cancellation before dispatch.

Catch the cancellation before it ships

Every box travels a 24-hour pre-shipping window. Some get flagged for cancellation mid-transit β€” press and hold one for a second to catch it before it hits the wall. Miss it, and it ships anyway.

€8 shipping + €35 goods = €43 lost on every flagged box that reaches the wall β€” illustrative numbers, not real ones.

Press and hold a flagged box to catch it.
0
Caught by hand
€0
0 boxes shipped before you caught them
β€”
Caught by the AI
β€”
Saved by the AI
πŸŒ™
πŸŒ™
πŸŒ™
πŸŒ™
πŸŒ™
πŸŒ™
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
β˜€οΈ
πŸŒ™
πŸŒ™
πŸŒ™
πŸŒ™

This is the everleakproof edge case from the real project: once a lingerie order ships, it's non-returnable, so the only cancellation window is before dispatch. A human has to notice a flagged order, decide, and click in time β€” the support bot reacted the instant a cancellation came in, catching a window a person only sometimes makes.


Work

TripleWhale Integration

dunlimited05/2020 – 06/2021

A fairly banal but necessary integration: synced Shipping Costs from the Shopify order logistics system into TripleWhale. Needed to get rid of manual data entry.


Work

AI-Character Telegram Bots

dunlimited05/2020 – 06/2021
Shut down β€” Telegram's inherently high cost-per-lead, combined with Stars/TON commission fees, killed the economics.
<10 sec
Deploy time

These bots sold access to scripted AI characters β€” a psychologist users cried talking to (voice messages via Whisper made that possible), and a fortune-teller taught to 'lie' convincingly: vague enough to always feel relevant, specific enough to feel true. Generated graphics and video kept it entertaining. The project shut down once the numbers stopped working: Telegram's cost-per-lead is inherently high, and monetizing through Telegram Stars meant a 50% commission converting to TON plus TON's own fee on top β€” combined, that killed the economics.

Built together with

Nikita Denisov β€” Computer Vision Engineer

Nikita built the computer-vision pipeline behind the generated graphics and video that kept these bots entertaining.

View on LinkedIn

Keep exploring