A high-performance, real-time crypto analytics and prediction platform.
Built in Rust with full data pipelines, multimodal ML, and MLOps for intraday tick prediction.
- โก Rust async pipelines (Tokio + Redis + Postgres) for low-latency ingestion
- ๐ Finnhub WebSocket streaming ~160k+ rows/day at <10ms latency
- ๐ Redis โ Postgres bridge for persistence of OHLCV & trades
- ๐งฎ DAG-based feature engine (10+ TA indicators) generating >100k training samples
- ๐ฐ Transformer-based sentiment analysis from Coindesk RSS, symbol-mapped
- ๐ค Multimodal signal fusion (price + volume + sentiment) via XGBoost + Transformers
- ๐ Grafana dashboards for real-time monitoring (latency, throughput, accuracy)
- โ๏ธ Northflank Blue-Green deployments with MLflow for experiment tracking & 99.9% uptime
| Component | Description |
|---|---|
โ
ws_ingestor.rs |
Connects to Finnhub WebSocket and streams live prices into Redis (<10ms) |
โ
fetcher.rs |
Periodically writes OHLCV from Redis into Postgres with TLS/NoTLS fallback |
โ
news_ingestor.rs |
Collects Coindesk RSS, maps to symbols, stores JSON headlines in Redis |
โ
dag_engine.rs |
Computes 10+ TA indicators (RSI, MACD, VWAP, etc.) for training datasets |
โ
xgboost_trainer.py |
Trains tick prediction classifier, logged via MLflow |
โ
sentiment_model.py |
Transformer-based sentiment scorer for financial news |
๐งช deployment/ |
Blue-Green deployment scripts with MLflow model registry + Northflank CI |
flowchart TD
A[Finnhub WebSocket] -->|Trades/OHLCV| B[websocket.rs]
B --> C[Redis]
C --> D[fetcher.rs]
D --> E[Postgres: stock_price_history]
C --> F[news_ingestor.rs]
F --> C
E --> G[dag_engine]
C --> G
G --> H[xgboost_trainer.py + sentiment_model.py]
H --> I[MLflow Registry]
I --> J[Northflank Blue-Green Deployment]
J --> K[Grafana Monitoring]
- ๐ฆ Rust (Tokio, async, Redis, Postgres)
- ๐ Python (XGBoost, Transformers, MLflow)
- ๐ Redis + PostgreSQL (real-time + persistence)
- ๐ Grafana (real-time monitoring & dashboards)
- โ๏ธ Northflank (Blue-Green deployment, CI/CD, 90% uptime)
- XGBoost tick classifier with DAG-engine features
- Transformer sentiment fusion for news-driven volatility
- Metrics tracked via MLflow: accuracy, F1, false-signal reduction (~18%)
- Continuous retraining + deployment with zero downtime
PRs are welcome!
If you're interested in real-time systems, crypto modeling, or ML infra, open an issue or contribute directly.
MIT ยฉ 2025 Swastik Nandy