SellerVai

Bangladeshi online sellers operate almost entirely through social media direct messages. A customer doesn't visit a website — they message the seller's Facebook page asking about availability and delivery. The seller manually reads every message, checks stock, replies with price and delivery info, collects the address, and confirms the order. For a seller handling 100+ messages a day during a promotion, this is physically unsustainable.

The second layer of complexity is reliability. Social commerce flash sales generate traffic spikes that are sudden and short. Meta's webhook delivery has a strict acknowledgment timeout, and if your handler is doing real work synchronously, you'll miss the window, trigger retry storms, and end up processing the same messages multiple times.

The ingestion layer was designed to do exactly one thing: verify the webhook signature and push the raw payload to a BullMQ queue. It returns 200 OK to Meta's servers in under 15 milliseconds. All actual processing — intent parsing, inventory lookup, response generation, message delivery — happens in worker processes that consume from the queue asynchronously.

Deduplication uses SHA-256 fingerprints of the channel + sender ID + platform message ID, stored in Redis with a 5-minute TTL using SET NX semantics. If a fingerprint already exists, the worker skips processing. This handles Meta's retry behavior cleanly without any application-layer coordination.

The bilingual parsing uses a two-tier approach. A regex engine extracts structured data — phone numbers, numeric sizes, city names, specific product codes — in under 2ms. An LLM handles intent classification for the parts that require understanding colloquial Bengali-Banglish phrasing.

SellerVai handles the full conversational sales workflow automatically. Incoming messages across all four platforms arrive in the queue, get deduplicated, classified by intent (stock check, price inquiry, order confirmation, COD request, delivery status), and routed to the appropriate response handler. The merchant sees a unified inbox with conversation history across channels.

COD orders — a significant fraud vector in Bangladeshi social commerce — go through an additional screening step that checks address completeness, order history, and behavioral signals before confirming.

During the first flash sale campaign after launch, traffic spiked to roughly 15x baseline over two hours. The queue absorbed the burst, webhook endpoints stayed responsive, and every message was processed exactly once. No duplicate orders, no missed confirmations.

Merchants using SellerVai eliminated the manual reply loop for routine inquiries. The bot handles stock checks, delivery queries, and COD confirmations without human involvement — escalating only messages that fall outside its confidence threshold.