February 20, 2026 · 13 min read

Designing a Safe AI-Assisted Trading System: From Stock Selection to Execution

AI Architecture Risk Management FastAPI Trading Infrastructure
[RESEARCH NOTE]
This article explains software system boundaries and stage-gated risk execution architectures. It is strictly for software engineering purposes.

1. Separating Prediction from Execution Authority

Building machine learning pipelines for automated decision systems requires strict isolation between probabilistic score generation and deterministic order execution.

2. Stage-Gated Selection & Risk Pipeline

Stock Selection & Risk Execution Boundary
flowchart TD
    A[NIFTY Stock Universe] --> B[Intelligent Stock Selection Service
0.25% Neutral Threshold Check]
    B --> C[Option Contract Enhancement
Upstox Real Lot Size Resolver]
    C --> D[Candidate Order Proposal]
    D --> E[Deterministic Risk Gate
Capital Cap & Session Check]
    E -->|Approved| F[Paper Trading Sandbox
Simulated 0.05% Slippage]
    E -->|Rejected| G[Audit Rejection Log]
    F --> H[PostgreSQL Trade History & PnL Ledger]

Diagram 1: End-to-end stage-gated execution pipeline.

3. Engineering Decisions & Tradeoffs

Decision Rationale Tradeoff
0.25% Market Neutrality Threshold Prevents dynamic option contract selection on low-conviction range-bound price moves. Filters out low-volatility trading opportunities.
Dynamic Upstox Instrument Registry Resolves true exchange lot sizes dynamically instead of relying on hardcoded defaults. Requires instrument token lookup before trade prep initialization.

4. References