ReplyBay
AI customer support for online stores. Email, WhatsApp, Instagram, Messenger, TikTok Shop and public comments land in one inbox, and each reply is drafted from the customer's order and the store's own policies.
- Role
- Co-founder
- When
- 2026 - now
- Links
- replybay.ai
The story
Small online stores drown in the same questions every day: where is my order, can I change the size, did you get my email. They arrive on Gmail, WhatsApp, Instagram and TikTok Shop, often at 2am, and every one is a small promise waiting to be kept.
I co-founded ReplyBay to give a solo founder or a five-person crew the kind of support desk that used to need a team. My co-founder brings the merchant side; I build the platform: the AI reply pipeline, the API and workers, both web apps and the infrastructure behind them.
The hard part is not writing a reply. It is knowing when a reply is safe to send. A wrong answer about a refund costs a store more than a slow one.
Constraints
- Replies go to real customers of real stores, so nothing unsafe may ever be sent.
- Customer data stays in the EU, and nobody's messages are used to train models.
- Every channel has its own API, limits and rules, from Meta to Shopify and TikTok Shop.
- Merchants in many countries, writing in their own language.
Decisions I made, and why.
Review first, then earn auto-send
Every AI reply is a draft until the store turns auto-send on. After that only replies above its confidence threshold, in categories it allows, go out. The rest wait, and an upset customer goes straight to a person.
Grounded in the order, with sources
Each draft is built from live order data and the store's own policies, then checked against them before it can be sent. The sources are shown next to the reply.
Safety as a test suite
The reply-safety contract lives in the tests: returns and refunds escalate, and nothing unsafe is sent. CI blocks any change that weakens it.
Temporal for every message
Each inbound message runs as a durable workflow: intake, context, draft, validation and routing. Provider outages and retries never lose or double-send a reply.
EU by design
GKE, PostgreSQL with point-in-time recovery, Redis and Temporal run in the Netherlands. Workload logging is off so message bodies never reach a global log bucket.
One platform, many channels
Gmail, WhatsApp, Instagram, Messenger, TikTok Shop, Shopify, WooCommerce and helpdesks like Zendesk, Gorgias and Re:amaze all feed one queue with the same rules.
The trade-off
Holding replies for review is slower than letting the AI answer everything. We chose it because trust is the product: a store turns automation up as the drafts prove themselves, and can always turn it back down.
What I built
The AI reply pipeline
Intake, context building, case memory, grounding, quality checks, validation and routing, each a separate tested step with a clear decision trace.
Comments and automations
Instagram and Facebook comments read like messages: a buying question gets a public answer, a comment-to-DM rule sends the link privately, and a complaint goes to a person.
Commerce context
Live order, tracking and product facts from Shopify, WooCommerce and TikTok Shop in every reply, so a reply about an order starts from the order.
Infrastructure
GKE in the EU with Terraform, Temporal, CloudNativePG and point-in-time backups. The Nuxt app and marketing site run on Cloudflare Pages.
Results
- Live product with a free plan, at replybay.ai.
- One inbox for email, DMs and public comments across five channels, plus helpdesks.
- Replies in 47 languages, with the app itself in seven.
- Core data and backups in Google Cloud's Netherlands region.
Stack
- FastAPI
- Temporal
- PostgreSQL
- Nuxt 4
- GKE
- Terraform
- OpenAI
- Cloudflare