AI Assistant & Intelligence Workflows
TRADEOS.tech includes a built-in AI assistant that lets you interact with the system in natural language. It is not a generic chatbot. It can be grounded in positions, signals, regime state, portfolio metrics, public-intelligence evidence, and review status.
The assistant operates in distinct modes, each with a different purpose and authority level.
Modes: Observer, Research, and Trader
Observer Mode
The Observer is a read-only analyst. It can answer any question about what the system is doing and why, but it cannot place or modify orders.
Personas available in Observer mode:
| Persona | Focus |
|---|---|
| Default Observer | General system status, open positions, recent signals, review status |
| Risk Manager | Drawdown exposure, circuit breaker state, risk limits, heat maps |
| Performance Analyst | Win rate, PnL attribution, per-signal-type breakdown, R-multiples |
| Regime Expert | Current HMM regime, BOCD change-point probability, regime history |
| Quant Coach | Signal quality, IC trends, calibration drift, what the system is learning |
| Research Reviewer | Thesis evidence, token identity, source coverage, publication readiness |
Each persona has a distinct analytical focus and system prompt tuned to its domain. Switching personas reshapes the assistant's perspective on the same underlying data — the Risk Manager looks at the same open position very differently than the Quant Coach does.
Research Mode
The Research assistant helps review public-intelligence drafts. It can explain why a symbol is on the thesis watchlist, what evidence supports the thesis, which uncertainty remains, what changed since the last checkpoint, and whether token identity includes chain and contract context.
Research Mode is read-only. It can help prepare review text, but it cannot publish, post, email external readers, or add unsupported claims.
Trader Mode
The Trader assistant has execution authority. It can propose and execute trades, adjust stops, close positions, or change configuration — subject to a confirmation flow before any action is taken.
Every trade action proposed by the Trader assistant requires explicit user confirmation before it is submitted to the execution layer. The system never acts unilaterally on a natural-language instruction.
Trade actions the Trader assistant can initiate:
- Open a new position (BUY / SELL, with symbol, size, order type)
- Close an existing position (full or partial)
- Adjust stop loss or take profit levels
- Override signal gates or circuit breakers (with operator authorization)
The Confirmation Flow
When the Trader assistant proposes a trade or action, a confirmation step is inserted before execution. The confirmation record carries:
- The proposed action in full (symbol, direction, size, order type)
- The reasoning the assistant provided
- An expiry timestamp — confirmations that are not acted on within their window are automatically discarded
Confirmation states: PENDING → CONFIRMED → EXECUTED or PENDING → REJECTED or PENDING → EXPIRED
This flow means you always have an opportunity to review and reject before anything hits the exchange. The assistant proposes; you decide.
Domain Knowledge
The assistant is grounded in curated knowledge about TRADEOS.tech's operational behavior — not generic trading theory. When you ask about a circuit breaker trip, the assistant understands how the DrawdownBreaker escalation works, how state persists across restarts, and what conditions trigger lockouts. When you ask about regime detection, it understands the HMM state space and how the BOCD override interacts with regime dwell requirements.
This domain knowledge is curated and maintained separately from the assistant's underlying language model, so it reflects the actual behavior of your running system rather than textbook definitions. For public-intelligence answers, source evidence remains the authority and the model only explains or rewrites what the evidence supports.
Live Context Integration
Every assistant query is enriched with live context from your system before it reaches the language model:
- Current symbol — if you are viewing a chart, the assistant knows which instrument
- Open positions — the assistant can see your current portfolio exposure
- Recent signals — the signal pipeline's recent output is available for discussion
- Regime state — current HMM classification and BOCD stability
- Portfolio metrics — equity, drawdown, session PnL
- Public-intelligence evidence — thesis status, chain and contract identity, source timestamps, material changes, and review cadence where available
This means you can ask questions like "why did the system not enter on that BTC signal?", "how exposed am I to high-volatility regime risk right now?", or "why is this token on the thesis watchlist?" and get answers grounded in your actual system state — not generic explanations.
Safety Constraints
The Trader assistant operates within the same risk envelope as the automated system:
- Proposed trades are subject to feasibility checks (capital availability, concentration limits)
- Circuit breakers still apply — if the DrawdownBreaker is tripped, the Trader assistant cannot override it without operator-level authorization
- All actions are logged to the audit trail with full lineage, including the natural-language instruction that initiated them
- Public-intelligence drafts remain subject to source, freshness, duplication, language, channel, dry-run, pause, and kill-switch policy gates
The assistant is a productivity interface, not a bypass. Trade actions go through the execution pipeline. Research actions go through the evidence and publisher policy pipeline.