The problem Merchant Mate solves

Screen Recording of our demo:

5. 
   Merchant Mate not only unifies sales and performance metrics from every delivery channel in real-time, but also transforms raw data into actionable insights pin-pointing peak hours, emerging menu favourites and hidden pain-points without a single manual spreadsheet. It continuously reads multilingual customer feedback, auto-classifies sentiment, and can launch the right remedy on its own: from drafting and sending escalation e-mails to support teams, to opening tickets and chasing follow-ups until the issue is closed. Day-to-day ops are equally friction-less one command updates menus, prices or branch details and the change ripples instantly across Swiggy, Zomato, Magicpin and any POS in the stack, wiping out the costly inconsistencies that lose orders. On the growth side, Merchant Mate is “crypto-ready”: through Coinbase Onchain AI it can spin up an ERC-20 loyalty token in seconds, fund airdrops to a curated customer list, or disburse on-chain rewards for rating milestones all without the merchant touching a wallet or writing Solidity. Promotions become programmable primitives: “roll out an extra 10 % weekend discount” or “reward my top-10 diners” triggers contract deployment, token minting and mass distribution in the background, while the Telegram concierge collects opt-ins and keeps the owner in the loop. By automating analytics, escalation workflows, omnichannel updates and even on-chain loyalty campaigns, Merchant Mate frees teams to focus on crafting memorable dining experiences and strategic expansion instead of routine coordination.

Challenges we ran into

When we first integrated a large language model, it routinely hallucinated facts—fabricating customer quotes or promotional details when our prompts lacked context—and delivered inconsistent or overly generic recommendations for ambiguous queries, all while adding a frustrating one‐to‐two‐second lag to each dashboard interaction. To address this, we built a retrieval‐augmented generation pipeline that pulls in real‐time sales figures, inventory data, and actual customer feedback before invoking the model, then wraps every request in a strict prompt template enriched with a couple of few‐shot examples. We dialed the temperature down near zero for fact‐driven tasks and layered in a lightweight validator that cross‐checks each generated claim against our database to catch any stray hallucinations.

Balancing speed and sophistication required a tiered approach: simple summary requests are handled by a tiny on-premise model that responds in under 200 ms, while complex, multi-platform campaign prompts are escalated to our full-sized cloud model. We pre-warm that heavyweight model shortly before predictable traffic peaks and cache its outputs so the UI can retrieve deep recommendations instantly. For longer narrative responses, we stream the first sentences immediately—giving users instant feedback—and progressively append the remainder as the model finishes, preserving both responsiveness and richness of insight.

Progress made before hackathon

None