Architecture

How RegimeForge works and why it should win Track 2

Live System Status

Loading system status...

What is RegimeForge?

A CoinMarketCap-native AI strategy skill for BNB Hack Track 2. Classifies crypto market regimes and generates structured, backtestable trading strategies.

How It Works

Data Ingestion

Consumes CMC Agent Hub market data — price, volume, RSI, funding rate, fear/greed, social volume, on-chain flows.

Signal Computation

Deterministic signals: momentum, sentiment, volatility, derivatives, on-chain (-1 to +1).

AI Regime Classification

NVIDIA NIM classifies market into 6 regimes: TREND_UP/DOWN, MEAN_REVERT_UP/DOWN, HIGH_VOL_BREAKOUT, CHOP.

Strategy Generation

Structured strategy spec with quantifiable entry/exit/invalidation rules, sizing, and horizon.

Validation & Critique

Deterministic validator checks backtestability. AI critique loop identifies weaknesses.

Backtest

Strategy applied to synthetic historical data. Returns drawdown, win rate, trade count, equity curve.

Explainability

Signal weights, reasoning, weak points, and thesis invalidators for every output.

Architecture

Modular TypeScript:

src/skill/cmc-skill.tsCore CMC Skill entry point

src/skill/prompts.tsStructured AI prompts

src/skill/parser.tsAI output parser + deterministic fallback

src/regime/classifiers.tsZod-validated schemas

src/regime/classifier.tsRegime classification

src/regime/features.tsSignal computation

src/regime/validators.tsRule validation

src/backtest/engine.tsBacktesting engine

src/backtest/metrics.tsTechnical indicators

src/backtest/scenarios.tsSynthetic data + presets

src/ai/nim-client.tsNVIDIA NIM client

src/orchestration/runner.tsPipeline orchestration

src/orchestration/critique-loop.tsAI critique loop

src/data/cmc-client.tsCMC data adapter

src/data/adapters.tsData transformers

src/ui/charts.tsxEquity curve + stat cards

src/ui/inspectors.tsExplainability builder

Why This Is More Than a Prompt Wrapper

✓ Deterministic signal pre-processing BEFORE AI inference
✓ Schema-validated outputs with zod
✓ Validator rejects vague strategies
✓ AI critique loop catches weak strategies
✓ Backtesting proves quantifiability
✓ Explainability with signal weights
✓ 6 distinct regimes
✓ Deterministic fallback when AI unavailable

Why This Should Win Track 2

✓ CMC Skill delivers backtestable strategy spec
✓ Strong CMC Agent Hub integration
✓ Real NVIDIA NIM AI inference
✓ Complete explainability
✓ Working backtest with equity curves
✓ Production-quality demo
✓ Comprehensive documentation