Why My Reorder Model Now Asks Hindsight Before Ordering

My forecasting system made a mistake for a retailer who strictly limited stock. The owner never wanted more than 35 cartons of noodles, but the model suggested ordering 50. That discrepancy between the numbers and the owner's decision is what I aimed to resolve with a new system called Hindsight. Hindsight is an open-source agent memory system that adds long-term business memory to the forecasting model.

The system is designed for small Indian retail shops, providing a dashboard with sales, profit, order count, and low-stock items. The system also includes billing options, a khata ledger for customer credit, an expenses ledger, and an advisor that can answer questions in Hinglish. The backend is a FastAPI service that acts as an orchestration layer, handling various stores of knowledge, including structured facts, machine learning forecasts, long-term business memory, and reasoning and conversation.

The key principle is that a forecast is a number, while a reorder decision is a number filtered through constraints that are not present in the sales table. These constraints include the owner's shelf limits, supplier relationships, and patterns in the owner's behavior. Instead of trying to anticipate these constraints in advance, I treated the owner's conversation as a valuable source of memory.

Every owner conversation and event that Hindsight detects is retained in a specialized "bank" for each store. When the system needs to make a reorder recommendation, it first gathers relevant data from the Postgres database and the forecast, then asks Hindsight what it knows that is relevant to the specific product. This allows the system to incorporate the owner's constraints and preferences into the decision-making process, leading to more accurate and context-specific reorder suggestions.