PROFIT

Inspiration

Store managers at convenience chains lose money in three quiet ways: shelves go empty when demand spikes, perishables expire, and slow products sit unsold. Most forecasting tools stop at "you'll sell 40 units tomorrow." We wanted to answer the next question: what should I do about it right now, and how much will it earn me? That became StockCopilot, an AI copilot that turns a demand forecast into restock orders and smart promotions a manager can understand, simulate and approve.

What it does

Forecasts demand per store × product with probabilistic bands (P50/P80/P90).
Detects buying behaviors: demand peaks, waste risk and products bought together.
Builds a commercial plan of up to 5 strategies per store (affinity combos, waste-rescue bundles, contextual offers) with start and end dates.
Shows the trade-off: for every promotion, the manager sees "Si no aceptas / Si aceptas" (without vs. with accepting) and can move a discount slider to explore the curve.
Suggests restock orders on a calendar, with suppliers ranked by on-time reliability.
Keeps a human in the loop: the agent proposes, and nothing is sent or activated without explicit confirmation.

How we built it

Data. We generated a synthetic but causally plausible dataset of 25 convenience stores in Saltillo: daily sales, inventory, expiry, suppliers, weather, local events, paydays and ticket lines.

Model. A CatBoost quantile model, trained as a SageMaker Training Job, predicts baseline demand. It minimizes the pinball loss for each quantile $\tau \in {0.5, 0.8, 0.9}$:

$$ L_\tau(y, \hat{y}) = \max\big(\tau\,(y - \hat{y}),\ (\tau - 1)\,(y - \hat{y})\big) $$

We validated it with four fixed 14-day time-based backtests.

Decision engine. Every strategy is scored by its incremental impact, the expected outcome with it minus the baseline without it:

$$ \Delta_{\text{net}} = \big(\text{margin}{\text{treated}} - \text{margin}{\text{baseline}}\big) + \text{waste avoided} - \text{stockout cost} $$

A strategy is rejected if it breaks a minimum margin, exceeds available inventory or creates an unacceptable stockout risk. The plan's value is the portfolio impact, not a naive sum of cards that might cannibalize each other.

Agent. An orchestrator built with LangChain/LangGraph on Amazon Bedrock reads the model outputs and explains them in plain Spanish. The ML models produce the numbers; the LLM only interprets them.

Infrastructure. API Gateway + Lambda (Go), DynamoDB for current state, S3 + Athena for history, EventBridge for scheduled jobs.

Frontend. Next.js, TypeScript, Tailwind, Recharts and Leaflet (sales heatmap). A typed data layer validates every response with zod and can switch between mock, hybrid and live API modes, so the UI team never had to wait on the backend.

Challenges

Parallel work in 24 hours. Five people were working on the model, the agent, the backend and the UI at once. We wrote an API contract first and built a deterministic mock backend with the same schemas, then switched endpoints to the real API one by one.
Keeping math out of the client. It was tempting to compute impact in the browser. We made it a hard rule that every financial figure comes from the backend, so the numbers on screen always match the model's.
Making AI trustworthy. An LLM that invents numbers is worse than none. We limited it to explaining engine outputs, made its chat responses a strict validated schema, and required human approval for every action.
Simple but honest UX. Showing uncertainty bands, portfolio effects and discount curves to a store manager in under five minutes meant hiding the technical details behind "¿Cómo se calculó?" (how was this calculated?) without hiding the truth.

What we learned

Forecasting is only half the product. The value is in turning a forecast into a decision with a price tag.
Measuring incremental impact against a baseline changes which promotions look good.
A typed contract plus a realistic mock is the best way to run a fast multi-team build.
Human-in-the-loop design builds trust: managers accept AI suggestions more easily when they can see what happens if they say no.