← Back to projects

Machine Learning

InvenIQ

Demand forecasting & inventory intelligence for quick-commerce

Sep – Oct 2025View source
Machine Learning
0.88
R² (variance)
91.4%
Stockout Acc.
22%
Cost cut

Overview

An end-to-end demand-forecasting and inventory system for Blinkit-style Indian dark stores — LightGBM forecasting wired into a multi-agent LangGraph pipeline and a real-time dashboard.

  • Trained on 5 years of Indian FMCG daily sales (913K rows) with 70+ engineered features — lag periods, rolling windows, EWM — and festival-aware post-processing for Diwali, IPL, Navratri and more.

  • Serves 10 dark stores and 50 SKUs with category-aware stocking windows (dairy 3-day, snacks 14-day, staples 30-day), safety-stock and EOQ optimization, and auto-resolving reorder alerts.

  • A three-agent LangGraph pipeline (Demand → Inventory → Logistics) explains every recommendation in plain language and drafts purchase orders for manager approval.

Tech stack

  • Python
  • LightGBM
  • Optuna
  • Pandas
  • Scikit-learn
  • LangGraph
  • Next.js
  • Groq
  • Gemini

Screens & results

InvenIQ landing — demand intelligence for quick-commerce, forecasting every SKU across the dark-store network.
InvenIQ landing — demand intelligence for quick-commerce, forecasting every SKU across the dark-store network.
Platform overview — LightGBM forecasting wired into a multi-agent LangGraph pipeline.
Platform overview — LightGBM forecasting wired into a multi-agent LangGraph pipeline.
Model scorecard — 88.3% R², MAE 3.41, RMSE 4.76 across 10 dark stores and 50 SKUs.
Model scorecard — 88.3% R², MAE 3.41, RMSE 4.76 across 10 dark stores and 50 SKUs.
Actual vs. predicted demand — LightGBM tracking daily sales with confidence bands.
Actual vs. predicted demand — LightGBM tracking daily sales with confidence bands.
Top engineered features driving the forecast (lags, rolling windows, calendar signals).
Top engineered features driving the forecast (lags, rolling windows, calendar signals).
Residual distribution — errors centred near zero, no systematic bias.
Residual distribution — errors centred near zero, no systematic bias.
Stockout-risk classifier confusion matrix — 91.4% accuracy.
Stockout-risk classifier confusion matrix — 91.4% accuracy.

Next project

Multi-Tenant WhatsApp Agent