Shahria EmonWhat happens in one conversation.
The buyer never typed a product name. MyMod worked it out from the photo, then answered like the shop's own owner.
The photo is the query
The buyer's image is embedded with a CLIP vision model and matched against the seller's catalog by visual similarity.
The variant is detected
The match resolves to a product, then to its option rows: color, size, per-variant stock and price. "Beige" maps to a real inventory row, not a guess.
The reply matches the buyer
The buyer wrote Banglish, so MyMod answers in Banglish, in a warm shopkeeper register. A Bangla buyer gets Bangla; an English buyer gets English.
The seller stays in control
The order is captured as pending. Anything MyMod was never taught goes to the seller instead of being invented.
A voice that never sounds like a bot.
Bangladeshi buyers spot an autoresponder instantly. The bar was replies that read like the seller's own thumb typed them.
The same shop gets "dam koto?", "price?" and "apu ei bag ta koto porbe" within the same hour. MyMod detects the script and register of each message and mirrors it: language, formality, even the short rhythm of Messenger commerce. Banglish was the hardest of the three, because it is not a language with rules but a habit with dialects, so the intent parser was built specifically for how Bangladeshi buyers romanize Bangla.
The second half of the voice problem is restraint. MyMod only states what the seller taught it. When a buyer asks something outside its knowledge, it hands off to the seller, and the question appears in a review queue so the seller can teach the answer once and never be asked again.
Matching a buyer's photo to the right variant.
F-commerce buyers shop with screenshots. For a seller that is memory. For software it meant a visual search engine over each shop's catalog.
Every catalog photo is embedded with CLIP into pgvector, and each photo is tagged to the option it shows, so a beige photo and a black photo of the same bag resolve to different variant rows. When a buyer sends an image, MyMod finds the nearest products, confirms with a vision model when the match is ambiguous, and answers with that variant's real price and stock. The same engine links the seller's Facebook posts to products automatically, so comment replies on any post quote the right numbers.
One person, the whole stack.
From the database schema to the Messenger webhooks to the logo, everything is a solo build.
A Next.js 15 TypeScript monorepo with Postgres 16 and pgvector for catalog search, Redis and BullMQ workers so webhook bursts never block replies, and a custom Facebook Graph and Messenger client with HMAC-verified webhooks, written directly against the platform. Gemini powers replies, order extraction and vision confirmation; a separate Python service runs CLIP image embeddings. Light and dark themes, and an onboarding a non-technical seller can finish alone.
The idea came from home.
My wife, Ritu, pointed at it: sellers in Bangladesh don't run stores, they run conversations.
Every product post collects dozens of comments asking price, stock and delivery, and every sale happens in Messenger. Before writing code we researched how sellers actually work and what they lose when messages go unanswered overnight. Buyers expect replies in minutes, write in three overlapping languages, and very often just send a photo: "eta ache?" Only after that validation did I start building.