Alpha Research Engine
Autonomous quant research for crypto, on local infrastructure.
Ingests market time-series and social sentiment, scores sentiment with local embeddings, computes rolling signal metrics, and backtests strategies with strict in-sample/out-of-sample separation. Optuna walk-forward tuning, LangGraph research agents, paper trading and a monitoring dashboard sit on top.
Skills demonstrated
Built on data science, delivered as AI engineering
Data science foundation
- Strict in-sample / out-of-sample separation
- Walk-forward tuning with Optuna
- Backtest metrics: Sharpe, Sortino, max drawdown
AI engineering
- LangGraph research agents
- Sentiment from local embeddings (Ollama)
- Autonomous optimizer behind a champion/challenger gate
- Built-in strategies
- 7
- Candles backfilled
- ~500k
- Tests
- 338
- IS / OOS split
- 2022–24 / 2025–
How it works
Design
Guardrails first
Strict in-sample (2022–2024) versus out-of-sample (2025 onward) separation, with no tuning on test data. Sentiment is timestamped at publication time, never ingestion time, so backtests can't see the future. Every pipeline stage is idempotent.
Sentiment without an LLM call
Each post is embedded once with a local nomic-embed-text model, and its score is the cosine similarity to bullish anchors minus the similarity to bearish anchors, clamped to [−1, +1]. It's deterministic, fast and free, and the embeddings double as searchable vectors in Qdrant.
Research automation
A parallel batch tuner tests every strategy × symbol combination across CPU cores, and a champion/challenger-gated optimizer plans and proposes new strategies. A rolling re-tune loop republishes parameters monthly, and paper trading forward-tests the tuned strategy on live data.
Stack