CePTFlow Intelligence Suite
CePTFlow was built to help retailers make stock decisions from their own sales history. It runs Holt Winters time series forecasting on historical sales, keeps the data in a structured SQL layer, and presents the forecasts through Plotly dashboards, with an AI assistant to help plan around them.
What I built
Demand Forecasting Engine
Demand forecasting using Holt-Winters Triple Exponential Smoothing to account for trends and seasonality.
Intelligence Vault (RAG)
A dedicated knowledge base that grounds the AI in actual sales documents, preventing hallucinations and ensuring factual accuracy.
Strategy Assistant
A conversational AI agent that analyses the "Vault" to provide instant answers on stock levels, peak periods, and inventory adjustments.
Multi-Modal Ingestion
Support for enterprise-grade CSV uploads, manual sales entry, and vision-based receipt scanning for "on-the-floor" data entry.
Cinematic UX
A premium "Charcoal & Gold" interface designed for high-end store environments, featuring smooth scroll-linked metrics and interactive Plotly visualizations.
How it works
Retrieval-Augmented Generation (RAG) Architecture
I architected a custom RAG pipeline that bridges the gap between structured SQL data and unstructured AI reasoning. When a query is initiated, the system performs a scoped retrieval of the most relevant sales history for that specific merchant. This context is then injected into the Google Gemini LLM, allowing the AI to cite specific dates and figures rather than providing generic business advice.
Mathematical Forecasting & Time-Series Analysis
Implemented Holt-Winters Triple Exponential Smoothing to handle three distinct components of business data: Level (baseline volume), Trend (growth/decline), and Seasonality (recurring patterns). This results in a forecast that adapts to the real-world complexity of retail cycles.
System Resilience & Self-Healing API Layer
Developed a Self-Healing Model Wrapper that monitors API health in real-time. If the primary AI model (e.g., Gemini 2.0) returns a 404 or 429 error, the system automatically discovers and rotates to the next most stable model available in the user’s specific region, ensuring a 99.9% success rate.
Result
Holt Winters forecasts fed by 3 ingestion modes: CSV, manual entry and receipt scanning
"Successfully resolved complex 429/404 API handshake issues through a custom self-healing model discovery algorithm, ensuring 99.9% uptime for the AI reasoning layer regardless of regional SDK versioning."