How Tachyon Works
Tachyon is not a single model making predictions. It’s a system of seven AI agents, each specialized, collaborating through a shared workspace.
Architecture
Binance API ──► PostgreSQL ──► Feature Engine (50+ features)
│
┌───────────────────┼───────────────────┐
▼ ▼ ▼
MM Claude MM Grok ML Model
(macro analysis) (sentiment scan) (gradient boosting)
│ │ │
└───────────────────┼───────────────────┘
▼
Signal Pipeline
(confidence-weighted vote)
│
▼
Final Prediction
(direction + confidence)
The Agents
MM Claude — Macro Market Mind. Analyzes fundamentals: on-chain metrics, exchange flows, macro indicators (DXY, yields, equities correlation), and institutional positioning. Runs every 3-4 hours.
MM Grok — Sentiment scanner. Reads X/Twitter in real-time for crypto sentiment, narrative shifts, and fear/greed signals. Produces a sentiment score and narrative summary.
ML Model — Gradient boosting classifier trained on 74,000+ hourly candles (2017-present). Uses only stationary features (returns, volatility ratios, oscillators — never raw price). Currently migrating from 5-class to 3-class for better accuracy.
Experia — ML researcher. Runs experiments, tunes hyperparameters, tests new features. Operates autonomously via Claude Code sessions.
Cody — Infrastructure. Builds pipelines, deploys models, maintains the system. Also autonomous via Claude Code.
Astra — Strategy and content. Designs experiments, writes research letters, manages public presence.
Tak — Coordinator. Makes architectural decisions, reviews proposals, manages resources.
Key Concepts
Regime Detection
Markets behave differently in different regimes (risk-on, risk-off, transition). A model trained on all data equally will underperform because the rules change. We’re building regime-aware models that adapt their predictions based on current market conditions.
The Autopsy Engine
When a prediction fails, the system runs an automated post-mortem:
- What did each agent predict?
- Where did the consensus break?
- Which features were most misleading?
- What would have been the correct signal in hindsight?
These autopsies feed back into model retraining and agent prompt improvements. The system literally learns from its mistakes.
Living Constitution
The agents operate under a shared set of principles that evolve over time:
- Never overfit to recent data
- Prefer calibrated uncertainty over false confidence
- Publish accuracy honestly, including failures
- The system’s code, prompts, and strategy are all modifiable by the agents themselves
Self-Improvement Loop
Tachyon agents can modify their own code, prompts, and infrastructure. Cody writes code. Experia runs experiments. Astra designs strategies. Tak reviews and approves. This creates a continuous improvement cycle where the system gets better without human intervention (founder approves major changes).
Data
- Source: Binance BTCUSDT, 1-hour candles
- History: 2017 to present (~74,000 samples)
- Features: 50+ engineered features (RSI, MACD, Bollinger bands, volume profiles, volatility measures, etc.)
- Holdout: Last 60 days always excluded from training for honest evaluation
- Storage: PostgreSQL (time series), Neo4j (knowledge graph)
What We Don’t Do
- We don’t use insider information
- We don’t trade yet (Phase 0)
- We don’t promise returns
- We don’t hide our failures