Project 01 • Production System

The
Baker.

Smart Retail Operations Platform

A production-grade full-stack application that bridges the gap between Customer Experience (CX) and Operational Efficiency. It solves the "Small Bakery Paradox": balancing demand volatility with the high cost of food waste.

Visit Live Site ↗View GitHub

The Problem

Small bakeries face two critical failures:

1. Customer Friction: Unsure if open, no inventory visibility, wasted trips.
2. Operational Waste: Over-production due to lack of data.

The Solution

A unified Hybrid Monolith platform:

Customer Portal: A Real-time PWA for live stock visibility and reservations.
Staff Dashboard: An AI-assisted decision support system recommending daily production levels.

The "Dual-Core" Prediction Logic

Core 1: Demand Model

Dynamic Regression

Objective: Predict Sales Velocity ("How many can we sell?").

y = β0 + β1(x1) + ... + βn(xn)
  • • Day of Week: One-hot encoded vectors (Saturday_Boost).
  • • Weather: Categorical mapping (Rain, Snow) to impact coefficients.
  • • Logic: Base Bias + Weighted Coefficients = Dynamic Target.
Core 2: Risk Model

Sparse Data Handling

Objective: Quantify Inventory Risk ("Probability of over-production?").

Challenge: Zero-Inflated Data (Waste events are rare)

Standard regression treats waste as outliers. My solution:

  • • Active-Day Filtering: Isolates days where production > 0.
  • • Risk Score: Calculates Average Loss per Active Day vs. Velocity.
  • • Analyst Note: Waste is operational variance, not linear weather function.

Tech Stack

  • Backend: Java 17, Spring Boot 3.2
  • Data Science: Python 3.x, Pandas, Scikit-Learn
  • Frontend: Thymeleaf (SSR) + TailwindCSS
  • Database: H2 (Dev) / PostgreSQL (Prod)
  • DevOps: GitHub Actions (CI) → Render (CD)

Key Capabilities

"What-If" Simulation Engine

Staff can manually override weather parameters (e.g., switch "Sunny" to "Rain"). Triggers in-memory vector recalculation (O(1)) without model retraining (O(N)).

Mathematical Transparency

Rejects "Black Box" predictions. The UI decomposes recommendations into constituent weights so staff trust the AI.
Prediction = Base (10) + Saturday (+2) + Rain (-1) = 11 Units

Automated Pipeline (CI/CD)

Just-in-Time Training: Python training script runs as a pre-build step in GitHub Actions. Ensures inference weights are always synchronized with the latest sales snapshot before deployment.