Lasanta India
B2B trade platform with grounded AI assistant
The AI assistant
A buyer-facing chatbot embedded directly in the marketplace and onboarding flow. Built on top of the live MongoDB catalog with strict grounding rules: every quote is indicative, every spec is sourced. Multilingual UX — English, Hindi, Telugu, and Arabic buyers each see product copy in their script.
The system prompt enforces honest behaviour — no fabricated stock, no invented staff contacts, low-confidence queries escalate to the admin queue with the buyer’s full context attached.
Real catalogue, real prices
Live products at /products. Real categories: metals & raw materials (Copper Millberry, Copper Berry, Steel HMS 1&2, Aluminum Scrap), edible oils (Crude / Refined Sunflower), beverages & wellness (Herbal Teas, Wellness Coffee). Per-MT pricing in USD scales with shipment size.
50+ global partners · 15+ countries · 1000+ completed trades.
Grounded dialogue
Sample buyer interaction — “Copper Millberry, price for 5 containers, CIF Rotterdam?” The assistant returns a 99.9% A- Grade Premium spec with an indicative CIF $6,400/MT subject to LME movement, offers SGS inspection, and queues a proforma. Sources cited: catalog rows + trade-doc templates.
Architecture
- Next.js 16 App Router on Vercel + a dedicated Conversion Studio worker on Google Cloud Run for heavy document compute. HMAC-signed internal API.
- Vercel AI SDK with Google (Gemini) + OpenAI providers behind a small router. System prompt + retrieval keep answers catalog-grounded.
- MongoDB Atlas (Mongoose) — users, products, orders, chat sessions, contacts, wallets, transactions.
- Razorpay billing with wallet, subscriptions, usage charges, plan-gated features.
- JWT + bcrypt + OTP-verified register/login/forgot-password, Resend transactional email.
What I learned
- Grounded beats generativein trade contexts. Buyers won’t tolerate hallucinated price quotes; they do trust a citation back into the catalog.
- Splitting heavy compute pays off. Keeping the user-facing app on Vercel while conversion runs on Cloud Run means cold starts stay low where users feel them.
- OTP onboarding works in B2B. Buyers expect identity verification; OTP + email is the lowest-friction path that still signals professionalism.
- Unicode in PDFs is a research project. Devanagari + Telugu + Arabic + RTL all interact. Font fallback chains had to be pinned and tested per template.