Standing Order

Standing Order is a purchasing agent for small shops. The owner texts it what they need ("Need 40 bags of flour for Friday"). It reads supplier prices from forwarded emails and text-based PDF catalogues, picks a supplier, asks the owner to approve, then places the order: by phone call through an ElevenLabs voice agent, or by email when a supplier has no phone. Built for Problem Statement 2, Memory you can govern. What an agent remembers is data someone has to answer for, so memory is an auditable part of the product: WHO CAN READ IT: every read and write goes through one tenant-scoped data layer, and Atlas Vector Search pre-filters by tenant. An automated test checks that one business never sees another's memories. HOW IT'S CORRECTED OR FORGOTTEN: an owner can fix a memory by text ("Flour from Gianella's Ingredients are $2.80 now"). That creates a new version and supersedes the old one. Owners can also expire a memory or forget it, which leaves an audit entry. Memories can carry an expiry and are purged by a TTL index. PROOF IT CHANGED BEHAVIOR: every order records the memories it used. In the demo, the order goes to the supplier chosen because of a remembered discount, and the dashboard's memory receipt shows that memory and where it came from. HOW IT WORKS: a text thread is many short sessions. A new session starts after silence or when a task completes, and memory carries over through a small context pack instead of the old transcript, which keeps token use low. Quotes are parsed once into structured data, and supplier choice and price math are plain code, not LLM calls. An order counts as confirmed only when the supplier's confirmation number is captured from the call transcript. STACK: MongoDB Atlas (data, Vector Search, TTL, LangGraph checkpoints), Voyage AI embeddings, LangGraph, DeepSeek, ElevenLabs Agents, Twilio (SMS and voice), Gmail IMAP for forwarded quotes.

Team