Data and stack
The production stack behind the consultancy: thirteen analytical platforms, thirty plus public data partners, and a deliberately boring engineering toolchain. Built once, maintained continuously, deployed for clients on demand.
The platform suite
The consultancy runs on thirteen production platforms, each one a working observatory or research toolkit rather than a slide deck. Argus is the macro-financial risk platform: 285 modules across six analytical layers covering bank stress testing, sovereign risk, financial stability indicators, and crisis early warning. Athena is the AI and compute economics observatory: four engines and roughly sixty collectors tracking model training costs, accelerator supply, datacenter buildout, and the diffusion of frontier AI into the broader economy. Aegis is the applied economics intelligence platform: 257 modules across eighteen analytical layers, with an AI brain wiring twenty two tools into research workflows.
Promethean is the renewable energy and transition observatory: four engines and over fifty collectors covering generation, capacity, emissions, and the political economy of decarbonization. Ceres is the food and agriculture observatory: production, prices, trade, climate stress, and food security across a comparable set of engines and collectors. Salus is the global health economics observatory: health expenditure, outcomes, systems capacity, and disease burden, drawing on WHO, World Bank, and IHME. Sisyphus is the long-run macro warehouse: a curated panel of fifty nine WDI indicators from 1960 to 2025. Hercules is its labor and human capital companion, with nineteen indicators on employment, education, and demographic structure.
TradeWeave is the trade intelligence platform: bilateral flows, tariff schedules, gravity estimation, and supply chain exposure. EconAI is the causal inference toolkit: difference in differences, synthetic control, instrumental variables, regression discontinuity, double machine learning. Delphi is the broader research toolkit: seventeen estimators, data ingestion pipelines, and an end to end paper generation pipeline. Axioma is the interactive economic theory reference: forty nine canonical theorems across trade, macro, micro, game theory, and finance, each with a working visualization. Strategos is the electoral intelligence platform: three hundred AI constituency agents tracking the Bangladesh 2026 cycle, extensible to other elections.
Public data partners
Every number on every chart traces to a primary public source. We do not invent data, and we do not buy proprietary feeds when a public series will do. The trade stack draws on CEPII (BACI bilateral flows and the Gravity dataset), UN Comtrade Plus for raw declarations, WITS for tariff and concordance work, USTR for U.S. trade policy filings, and USITC for tariff schedules and Section 232 and 301 actions.
Macroeconomic and financial work pulls from the IMF (World Economic Outlook, Balance of Payments, Financial Soundness Indicators, Government Finance Statistics), the World Bank (World Development Indicators, Health Nutrition and Population, Global Financial Development Database), and the OECD (STAN industrial database, TiVA trade in value added). Domestic U.S. macro and labor data come from BLS, BEA, and FRED, with FERC and NERC for electricity reliability and EIA for energy supply, prices, and consumption.
Food and agriculture run on FAOSTAT for production, trade, and food balance sheets. Health work draws on WHO (Global Health Observatory, Global Health Expenditure Database) and IHME Global Burden of Disease for cause specific mortality and DALY estimates. Energy and emissions analysis combines IRENA for renewable capacity, IEA for balances and investment, Ember for power sector tracking, OWID for clean curated time series, and EDGAR for emissions inventories.
The AI and compute stack relies on MLPerf benchmarks, Epoch AI's training compute and model database, and SemiAnalysis for accelerator and datacenter intelligence. Across roughly thirty public sources, the operating rule is the same: every series is versioned, vintage stamped, and re fetchable from the original endpoint.
Technical stack
The backend is deliberately conservative. Python 3.11 across all observatories, with FastAPI and Uvicorn for HTTP, aiosqlite for storage (SQLite chosen for portability and zero ops, not for scale theater), httpx for async collection, and APScheduler for cron. Statistical work runs on statsmodels, linearmodels, and scikit-learn, with econml and doubleml for causal machine learning, rdrobust for regression discontinuity, and pandas plus numpy underneath everything.
The front end is Next.js 16 on React 19 with TypeScript throughout. Visualization is split by job: D3.js v7 for bespoke charts, Deck.gl v9 layered on MapLibre for geospatial work, Plotly for quick interactive exploration, and KaTeX for inline mathematics. The design system is dark first, with a small palette and three typefaces (Inter, JetBrains Mono, Source Serif 4).
AI integrations run through the Anthropic SDK for production agent workflows and Hugging Face for open model access and dataset retrieval. Infrastructure is intentionally minimal: uv for Python dependency resolution, systemd units for service supervision, Cloudflare for DNS and edge caching, and Let's Encrypt for TLS. Deployment is rsync plus a systemd reload, scripted per project. No Kubernetes, no service mesh, no managed data warehouse. The point is that a single engineer can understand, deploy, and debug the entire stack end to end.
How we extend the stack for clients
Most engagements use the existing platforms as is, with the consultancy doing the analysis and delivering findings. When a client wants the platform itself, we offer two installation patterns. White label hosted: we run a dedicated instance on our infrastructure, branded to the client, with their data and access controls, billed as a monthly subscription that covers compute, storage, collector maintenance, and indicator updates. Self hosted: we install the platform on the client's own VPS or cloud account, transfer credentials, and document operations, with a fixed implementation fee plus an optional annual maintenance retainer.
Knowledge transfer is part of every install. The client team gets a working repo, a runbook, and direct access to the engineer who built the system. Maintenance retainers cover collector breakage when upstream APIs change, indicator additions, security patches, and quarterly model refreshes. Clients who want deeper customization (new collectors, bespoke dashboards, integration with internal data) can scope that work separately or roll it into the retainer. Source for EconAI, parts of TradeWeave, and the observatory templates lives on github.com/deluair under permissive licenses, so clients can audit and extend the code directly.