Work

Inuka Risk Radar

Kafka
PostgreSQL
FastAPI
Next.js
XGBoost
n8n

Predictive intelligence and program automation platform that flags at-risk beneficiaries and triggers field-worker outreach in real time.

Inuka Risk Radar Dashboard

Predictive Intelligence & Program Automation for the Inuka Foundation

Inuka Risk Radar is a real-time risk-scoring and automation platform that identifies program beneficiaries (students/participants) at risk of dropping out, based on field telemetry like attendance, assignment completion, and travel distance. When a beneficiary's risk score crosses a threshold, the system automatically triggers field-worker outreach — all visualized on a live, offline-capable dashboard.

How it works

  1. Ingestion — Field events stream through Kafka (Redpanda) onto a beneficiary.telemetry topic.
  2. Storage — A Kafka consumer lands raw events in PostgreSQL (Bronze), which is joined against demographic/reference data into a deduplicated beneficiary identity graph (Silver), then aggregated into regional stats and forecasts (Gold).
  3. Intelligence — An XGBoost/Scikit-Learn model scores beneficiaries on request via POST /api/v1/evaluate, returning a risk tier and the human-readable drivers behind the score.
  4. Automation — Scores above the risk threshold trigger an n8n workflow, which notifies the assigned field worker (SMS via Twilio, or email as a demo-safe fallback).
  5. Experience — A Next.js dashboard consumes all of this via REST and Server-Sent Events, with an offline-first IndexedDB queue so field data entered without connectivity syncs back once reconnected.

Technologies

  • Apache Kafka (Redpanda)
  • PostgreSQL
  • Scikit-Learn / XGBoost
  • FastAPI, SQLAlchemy, Uvicorn
  • n8n (self-hosted)
  • Next.js (App Router), Tailwind, Shadcn UI, MapLibre GL
  • Docker Compose, Caddy, Cloudflare Tunnel

Deployment

The full stack (Kafka, Postgres, FastAPI backend, Next.js frontend, n8n) runs via Docker Compose on Oracle Cloud. Caddy reverse-proxies the frontend and backend behind a single public subdomain, and a Cloudflare Tunnel exposes it without any inbound ports — no public IP or firewall configuration needed.

Data & privacy

All data used — training data, demo data, and anything shown during judging — is synthetically generated. No real personal or beneficiary data is processed or stored.

Where To Find The Resource