Last-mile Early Warning System

Forecasts delivery stress 30 minutes ahead and lets an AI agent explain it

PythonXGBoostLangGraphClickHousePostgreSQLSlack

A production early-warning system for last-mile delivery: it tracks ~21 KPIs every 2 minutes, forecasts 15 and 30 minutes ahead, and hands anomalies to a LangGraph agent that finds the root cause and alerts the ops team on Slack.

The problem

Delivery operations only noticed demand spikes and rider shortages after orders were already late. They needed warning early enough to act.

Architecture

ClickHouse event data → 2-minute KPI slots stored in PostgreSQL
Anomaly detection with z-scores against day-of-week and rolling baselines
XGBoost forecasts at 15- and 30-minute horizons drive a stress score
LangGraph agent with long-term memory for root-cause analysis and a mitigation ladder
Tiered L1–L3 alerts to Slack, plus simulation and velocity models running alongside

Results

~21 KPIs scored every 2 minutes
15 & 30-minute forecasts
Agent-written root cause in every alert

More projects

Healthcare Document AI

Healthcare data

RAG over pharmacy-benefit reports: extraction, summaries and Q&A

PythonRAGpgvectorAzure OpenAIDocker

Legal AI Content Platform

Legal services

Production website + AI publishing pipeline for a law firm

Next.jsTypeScriptPostgreSQLDockerLLMs

Industrial AI Platform

Mining, food & process industries

Secure, tested full-stack platform for industrial process optimisation

Node.jsReactMongoDBAWSVault

All projects