Decentrathon 4.0, GeoOS

Project Overview

GeoOS is a geospatial analytics platform built in 48 hours as a solo developer (Team RPR) for inDrive's Case 2 at Decentrathon 4.0, the largest hackathon in Kazakhstan. The platform provides interactive map visualizations with AI-powered anomaly detection for ride-hailing data analysis. Placed 7th out of 15+ competing teams, scoring 83 points.

The Challenge

inDrive needed a way to identify unusual patterns in their geospatial ride data, demand anomalies, pricing outliers, and route irregularities across Central Asian cities. The challenge required:

• Processing large geospatial datasets in the browser
• Real-time anomaly detection without server-side ML infrastructure
• Interactive map visualizations that non-technical stakeholders could use
• Building a complete, demo-ready product in under 48 hours as a single developer

The Solution

Built a full-stack geospatial analytics dashboard featuring:

• Interactive 3D Maps, deck.gl layers with hexagonal heatmaps, arc visualizations, and scatterplot overlays rendered on MapLibre base maps
• Client-Side ML: TensorFlow.js anomaly detection running entirely in the browser, detecting demand spikes and pricing outliers in real-time
• Analytics Dashboard: Recharts-powered charts showing temporal trends, distribution analysis, and anomaly breakdowns
• Filtering System: Time range, geographic bounds, anomaly type, and confidence threshold filters

Technical Implementation

Frontend Architecture

• Next.js 14 with App Router for the application shell
• deck.gl for WebGL-accelerated geospatial rendering (hexagon layers, arc layers, scatterplot layers)
• MapLibre GL JS as the base map renderer (open-source, no API key required)
• Recharts for statistical visualizations alongside the map

AI/ML Pipeline

• TensorFlow.js for in-browser anomaly detection
• Isolation Forest algorithm adapted for geospatial point data
• Z-score statistical analysis for demand spike detection
• Sliding window approach for temporal anomaly identification

Data Processing

• GeoJSON parsing and transformation pipeline
• Spatial indexing for efficient geographic queries
• Web Workers for heavy computation without blocking the UI thread

Impact & Results

• 7th place out of 15+ teams at Kazakhstan's largest hackathon
• 83 points scored by judges (technical implementation + business value + presentation)
• Solo entry, competed against teams of 3-5 developers
• Fully functional demo delivered within the 48-hour constraint
• Live deployment at geo-os.vercel.app immediately after the event

Lessons Learned

• Scope ruthlessly under time pressure: I planned features for a 5-person team, then cut 60% before writing any code. The features I shipped were polished; the ones I cut would have been half-broken. Scoping down was the single best decision.
• Client-side ML is viable for demos but not production: TensorFlow.js ran the anomaly detection smoothly in Chrome on a MacBook, but performance varied wildly across browsers and hardware. A real product would need a server-side inference endpoint.
• deck.gl has a steep learning curve but enormous payoff. The first 6 hours were frustrating (coordinate systems, layer lifecycle, viewport syncing). Once it clicked, I could add new visualization layers in minutes. Worth the upfront investment.
• Hackathons reward storytelling as much as code: Teams that placed above me had simpler tech but better narratives. The judges cared about "why does this matter for inDrive?" more than "how does the isolation forest work?"
• Solo competing builds confidence: Knowing I placed 7th alone against full teams proved I could deliver end-to-end under extreme pressure.