GeoPulse — Geopolitical Intelligence Platform

Geopolitical events (military conflicts, sanctions, trade disputes, diplomatic shifts) have immediate market impact, but existing tools are siloed — ...

Financial analysts manually monitored dozens of news sources to assess geopolitical risk — no automated ML pipeline existed to ingest articles,...

No interactive globe visualisation correlated geographic events with market movements — analysts couldn't click a country to see region-specific...

Existing geopolitical risk tools lacked what-if scenario simulation — no way to model macro shocks (oil supply disruption, interest rate changes,...

AI trading signals in the market lacked transparency — no causal reasoning chains explaining why a signal was generated, no event-to-market reaction...

No narrative detection to identify emerging geopolitical themes — analysts couldn't see which stories were clustering (e.g. 'South China Sea...

Separated backend into 9 routers, 12 services, and 8 models — each service encapsulates one ML pipeline stage or data domain, routers handle HTTP...

APScheduler for automated ML pipelines over Celery — lighter weight for scheduled jobs (no broker dependency), sufficient for 6 periodic tasks at...

DistilBERT for sentiment over full BERT or GPT — 60% smaller, 60% faster, retains 97% of BERT's accuracy. For real-time sentiment scoring every 10...

BART-MNLI zero-shot classification over fine-tuned classifier — no labelled geopolitical event dataset required. Zero-shot handles new event...

LightGBM for macro predictions over deep learning — trains in seconds on tabular features (sentiment scores, event counts, GTI values), handles...

BERTopic with Sentence Transformers for narrative detection — produces human-readable topic labels from article clusters, handles dynamic topic...

globe.gl + Three.js for 3D visualisation over Mapbox Globe or Cesium — globe.gl provides high-level abstractions for points, arcs, and polygons on a...

react-simple-maps for 2D choropleth over Mapbox/Leaflet — lighter weight (no tile server dependency), D3-geo projection built-in, sufficient for...

lightweight-charts for candlestick display over TradingView widget — open-source (no API key), lightweight (~40KB), financial-grade OHLC rendering...

Rule-based signal engine over ML signal generation — deterministic signals with explicit reasoning chains (users can audit why a signal was...

What-if scenario engine with linear parameter adjustments — simple slider-to-risk mapping enables rapid iteration. Full Monte Carlo simulation would...

Next.js 14 with App Router over Pages Router — leverages React Server Components for initial data loading, streaming for progressive UI hydration,...

DistilBERT sentiment is general-purpose — not fine-tuned on financial/geopolitical text. Phrases like 'aggressive sanctions' score negatively despite...

Zero-shot classification into 9 categories has lower precision than a fine-tuned multi-class classifier — ambiguous events (e.g. 'trade sanctions'...

GTI recalculation every 15 minutes creates a lag window — events occurring between recalculations aren't reflected in risk indices until the next...

LightGBM predictions depend on feature quality — garbage-in-garbage-out from upstream sentiment/classification errors propagate to macro predictions....

BERTopic narrative detection requires sufficient article volume — fewer than ~20 articles in a time window produces unstable topic clusters. During...

globe.gl 3D rendering is CPU/GPU intensive — older devices or browsers without WebGL2 support may experience frame drops. The 2D map tab serves as a...

No real-time WebSocket push — frontend polls APIs at intervals rather than receiving server-pushed updates. WebSocket/SSE would reduce latency for...

Would improve: Fine-tune DistilBERT on labelled geopolitical sentiment corpus, add event-driven GTI recalculation, implement WebSocket for real-time...

Production-grade geopolitical intelligence platform unifying news ingestion, ML analysis, risk scoring, trading signals, and interactive...

3D globe (globe.gl + Three.js) with country risk polygons, severity-based altitude extrusion, night-earth texture, and real-time arc animations for 4...

2D choropleth map (react-simple-maps) with per-country financial analysis: region-aware asset selector, live candlestick chart (lightweight-charts...

AI trading signals with full trade setup (entry, stop loss, target, R:R, ATR, position sizing), causal reasoning chains with contribution bars,...

Automated ML pipeline with 6 scheduled jobs: RSS news ingestion (10 min), DistilBERT sentiment + BART-MNLI classification (10 min), GTI recalculation...

25+ REST API endpoints across 9 routers with clean separation: health, news (ingest + query), events (CRUD + ML processing), risk (GTI global +...

What-if scenario engine with 5 macro sliders (Oil Shock, Rate Change, Escalation, Supply Chain, Cyber Threat) adjusting regional risk indices and...

GitHub Actions CI/CD: frontend job (npm ci → next lint → next build), backend job (pip install → syntax check → module import validation for 9...

Docker Compose full-stack deployment (frontend + backend + PostgreSQL) and Makefile with 16 commands (install, setup, dev, dev-backend, dev-frontend,...

Frontend: Next.js 14, TypeScript, TailwindCSS, Framer Motion

3D Visualisation: Three.js, globe.gl (country polygons, arc animations, altitude extrusion)

2D Map: react-simple-maps (D3-geo choropleth with click interactions)

Charts: lightweight-charts (financial-grade OHLC candlestick rendering)

Backend: Python, FastAPI, Uvicorn, SQLAlchemy (async), APScheduler

NLP: HuggingFace Transformers (DistilBERT sentiment, BART-MNLI zero-shot classification)

ML: LightGBM (macro risk predictions), BERTopic + Sentence Transformers (narrative detection)

Database: PostgreSQL 16 (Supabase-ready), async SQLAlchemy ORM with 8 models

Infrastructure: Docker, Docker Compose, GitHub Actions CI/CD, Makefile (16 commands)