Skip to content
varun.rout

Jump to

Search pages, projects and repos

← Back to work

Energy & climate · live repo

Live repoReal data

Climate Transition Risk Intelligence Platform

Independent research · real Azure production deployment

View the repo →

What this is

Country-level decarbonisation and transition-risk analytics on public data (Our World in Data, World Bank), with a real Azure production deployment and a published, honestly-undershot coverage finding.

Metrics

Backtest reproduction

0.0262 MAE

vs documented scratch finding 0.0263

Interval coverage (honesty finding)

76.3% measured

vs 90% target, 114 splits

Energy-feature gate

p ≤ 0.10, robust at ±10/20/30% weight perturbation

Reproducibility (v1.0.0)

287 Python tests + 31 frontend tests, independently re-run clean-checkout

Azure footprint

8 resources, independently confirmed via Azure MCP

Stack

PythonTerraformAzure Container AppsADLS Gen2GitHub Actions CIReact/TypeScriptFastAPI

Provenance

Solo-authored, public, CI-green. A real Azure production deployment (Terraform-managed, least-privilege managed identities, weekly schedule) runs the full pipeline end to end against live storage. The 76.3% interval-coverage figure is reported as an open undercoverage finding, not tuned away, in the same style as the negative results elsewhere in this portfolio.

What I would not claim

  • a completed Power BI/PBIX native report (superseded by the live React/TypeScript dashboard)
  • real-time, streaming or low-latency serving
  • regime-aware or recency-weighted forecasting in production (evaluated, explicitly not promoted)
  • external or paying customers
  • unqualified "production-grade"

Does it earn the deployment?

A rolling-origin backtest reproduction, the coverage shortfall published rather than tuned away, and the pre-registered gate an energy feature had to clear before it reached production.

6 origins (2010–2022), 19 G20 countries. The bootstrap model's median absolute error closely reproduces a documented scratch finding; its interval coverage does not — and that miss is reported, not tuned away.

Published miss, not tuned away90% interval coverage

76.3%

measured, across 114 splits

90%

target

90% prediction-interval coverage measured at 76.3% across 114 splits — an open undercoverage finding, published as-is.

Data provenance. Committed backtest and gate outputs from climate-transition-risk-platform, generated into data/climate.json. Public data (Our World in Data, World Bank). Real Azure production deployment, independently verified.