Cynosure

The problem

Trading systems usually mean cloud LLM calls, data leaving the machine, and a "strategy" that's really just the model improvising. For perpetual swaps on OKX I wanted the opposite: a system that runs 24/7 fully locally, no cloud dependency, and, more importantly, one where the LLM is a synthesizer, not a reasoner. All quantitative work happens in deterministic Python; the model's job is to read a brief and write a thesis.

The hardware constraint that shapes everything: an RTX 4070, 8GB VRAM, running alongside everything else.

The approach

Every 15 minutes, per ticker (crypto majors, gold/silver, equity index perps), an expert pipeline compresses raw market data into a ~500-token MarketBrief:

┌─────────────────────────────────────────────────────────────┐
│  EXPERT PIPELINE (pure python, deterministic)               │
│  technicals  : EMA9/20/50 · RSI · MACD · bollinger · ATR    │
│  flow        : OFI / VPIN · vwap deviation · toxicity       │
│  forecaster  : timesfm 2.5 (200M) · ±0.30% threshold        │
│  orderbook   : 20-level L2 · imbalance · whale walls        │
│  regime      : fear & greed · btc regime · 24h news         │
└──────────────────────────┬──────────────────────────────────┘
                           ▼
        qwen3.5:4b via ollama, synthesizer, /nothink, temp 0.1
                           ▼
        DETERMINISTIC RISK ENGINE, the llm has no vote here
                           ▼
                   okx order · sqlite journal

• Experts over raw candles. The LLM never sees candle arrays, it sees the brief. Hallucinated signal counts get replaced with Python-computed values before risk gates run.
• Deterministic risk engine. Daily loss limits, circuit breakers, margin guard, correlation check (≥0.85), an EV gate (≥0.5%), half-Kelly sizing, ATR-adaptive OCO stops. The README line holds in code: the LLM has no role in risk decisions.
• Persistence gates. A BUY is blocked after a single-cycle bullish shift ("one cycle is noise"); SELLs after entry are left to OCO.
• Devil's advocate. Borderline BUYs (confidence 0.65–0.78) trigger a second LLM call that can return ABORT or REDUCE_SIZE.

Serving on a budget

Qwen3.5-4B runs through Ollama with thinking disabled (/nothink + think=false) at temperature 0.1, 5–8 seconds per call on the 4070, a full 15-ticker cycle in about 2 minutes. TimesFM 2.5 (200M) sits alongside for zero-shot directional forecasts. A custom MCP server wraps the OKX API with 11 tools.

Results and what I learned

• "LLM as synthesizer" beats "LLM as reasoner" in domains with rich deterministic pre-processing: the model's failure modes (invented counts, confident nonsense) get filtered before they can matter.
• Deterministic risk gates are non-negotiable: every check that matters is plain Python with no model in the loop.
• Local-only is a feature, not a compromise, no cloud latency in a 15-minute loop, and market data never leaves the network it belongs to.
• Paper mode for weeks, always. The system ships with disclaimers in its config.