Back to portfolio

♻️

Eco Drop

AI-Powered E-Waste Collection Scheduler

A full-stack, AI-powered application that manages e-waste collection scheduling — users book pickups while administrators optimise routes, demand and carbon impact with machine learning.

Overview

Eco Drop pairs a modern React frontend with a FastAPI AI backend. It forecasts pickup demand, clusters collection hotspots, optimises routes and quantifies the carbon saved by each optimisation.

Problem Statement

E-waste collection operators schedule pickups reactively, driving inefficient routes with half-empty vehicles — wasting fuel, time and emissions budget.

Architecture

  • Vite + React + TypeScript client with shadcn-ui and Tailwind CSS
  • Leaflet map layer and Recharts analytics dashboard
  • Supabase for auth and persistence
  • FastAPI AI service for prediction, clustering and routing
  • Ollama + LLaMA3 for strategy recommendations
  • OpenCage geocoding for address resolution

Workflow

  1. 01User schedules a pickup with an address, resolved through geocoding
  2. 02Demand model forecasts volume per zone
  3. 03K-Means clusters requests into collection hotspots
  4. 04Nearest-Neighbour routing produces an optimised vehicle route
  5. 05Carbon calculator scores emissions saved and the LLM drafts a strategy note

Features

  • Demand prediction with a custom ML model
  • Hotspot clustering via K-Means pickup detection
  • Route optimisation using a Nearest Neighbour algorithm
  • Carbon optimisation with an emission calculator
  • LLM strategy recommendations powered by Ollama + LLaMA3
  • Geocoding through the OpenCage API

Technology Stack

Vite + ReactTypeScriptTailwind CSSshadcn-uiLeafletRechartsSupabaseFastAPIK-MeansOllama + LLaMA3

Challenges & Solutions

Raw pickup addresses were unusable for routing.

SOLUTION · Added an OpenCage geocoding pipeline with caching and fallback matching.

Naive routing scaled badly as requests grew.

SOLUTION · Clustered requests with K-Means first, then ran Nearest Neighbour inside each cluster.

Operators needed the 'why', not just the route.

SOLUTION · LLaMA3 generates a short strategy summary alongside every optimised plan.

Results

  • Shorter routes per collection run through cluster-then-route optimisation
  • Quantified carbon savings surfaced directly in the admin dashboard
  • End-to-end flow from citizen booking to optimised operator route