Ask about the LLM Advisor workflow, telemetry, risk controls, or how signals map to actions. I use docs and Supabase metrics.
LLM Advisor
Paper-trading ops dashboard with live decision ledger, win-rate tracking, and LLM veto logs (OpenAI gpt-5.4-nano). LLM adds context and adjusts thresholds; rules keep control. Alpaca paper account, Supabase telemetry, real-time heartbeat monitoring.
PythonOpenAI APIAlpacaPandasBacktesting
USER: Analyze SPY sentiment
AGENT: Volatility high. Reducing position size.
Ask the Data
This assistant answers from LLM Advisor documentation and Supabase telemetry. Ask about run health, risk controls, signal flow, or how live metrics relate to the write-ups on this page.
LLM Advisor Chat
Assistant: project docs plus live telemetry for run health, risk controls, and trading loop behavior.
System Design
1. LLM Context Layer
OpenAI gpt-5.4-nano runs market analysis periodically (every 15 minutes by default) and returns threshold multipliers and confidence signals. The LLM adds context but does not bypass hard risk limits.
2. Rule-Bound Execution
The execution engine computes rolling mu/sigma/z-score states and evaluates MR/TC setups against configured thresholds before sending bracketed orders through Alpaca paper account.
Live Monitoring Dashboard
Live ops dashboard
Paper equity, win rate, trade breakdowns by symbol/setup, decision ledger with LLM vetoes, and execution funnel (signals → validation → orders) live at the standalone ops dashboard. Heartbeat monitoring shows real-time agent health and telemetry freshness.
Open the live ops dashboardThis dashboard is fed by the telemetry API. Production serves Supabase telemetry only, while local file fallback is reserved for non-production debugging.
Data source: No telemetry
Telemetry as-of: N/A
Last generated: N/A
Anchor date: N/A
Loading telemetry...
Coverage window: 0 run day(s)
Trades indexed: 0
Order events indexed: 0
Runtime loops seen: N/A
Feed mode: Unknown
Last Price Update
N/A
N/A
P/L Windows
N/A (7d)
1d: N/A | 30d: N/A
Recent Trades
0
Count in last 7 days of sample window.
Success Rate
N/A
0 trades available.
Risk/Reward
N/A
Avg win: N/A | Avg loss: N/A
Agent State
NO FEED
Loop: N/A | Symbols: N/A
Signal Flow
0
Rejected: 0 | Parser errors: 0
Order Attempts
0
Filled: 0 | Failed: 0
Recent Trades
No trades found in telemetry yet.
Recent Execution Events
No signal or order lifecycle events found yet.
Model Visuals
Feature-importance plots from current training artifacts.
SPY feature importance
QQQ feature importance
IWM feature importance
Automated Risk Manager
The runtime configuration enforces strict limits before execution: bounded risk per trade, minimum reward/risk, fixed session windows, and end-of-day flattening.
src/core/config.py + config/thresholds.py
# Runtime defaults (overridable via env) max_risk_per_trade_percent = 1.0 min_risk_reward_ratio = 1.5 trading_window_start = "09:30" trading_window_end = "12:00" end_of_day_close_time = "15:50" # Base STDEV thresholds mr_arm_z = 1.2 mr_trigger_z = 0.6 tc_arm_z = 1.8 tc_trigger_z = 0.6 atr_multiplier_sl = 1.4 atr_percentile_cap = 85.0
Notebook Analyses
Week-2 notebook work is now scaffolded to publish reproducible analysis artifacts for this project.
Offline simulation snapshot
Versioned JSON under public/data/llm_advisor_backtest_snapshot.json — reproducible headline stats for portfolio / LinkedIn, with explicit limitations (technical replay only; no LLM overlay in this batch).
Data Caveat
Metrics shown here are evidence-backed but mixed-source. If the feed mode is Backtest stream, P/L and win-rate reflect historical simulation artifacts; if it is Live stream, values come from persisted runtime telemetry. Deep-dive claims are restricted to what is currently materialized in those sources.