Trading AI Assistance — Backend Architecture

The backend for a trading-assistance platform whose specification lists 19 AI capabilities and 19 software capabilities — indicators, strategy execution, chart-image analysis, explainable signals, sentiment and news impact, risk management, backtesting and alerts. The repository turns that specification into explicit engineering boundaries and a working foundation.

The non-negotiable rules are written into the architecture: a trading signal is never presented as guaranteed profit; AI output is an analysis, not a command; every prediction must be traceable to model version, input snapshot, strategy configuration and timestamp; and backtesting must prevent look-ahead bias and data leakage.

A modular monolith with explicit domain boundaries and asynchronous workers, in a pnpm workspace: a NestJS HTTP API (/api/v1), a BullMQ worker, a framework-free domain package, an application package of use cases and ports, an infrastructure package (Prisma, Redis, BullMQ, logging adapters), typed configuration, and shared contracts for envelopes, error codes and queue names.

ML workloads are placed in a separate Python inference/training service where Python-native tooling is required. The documented AI architecture covers a model registry with model versions, inference requests and evaluation metrics, explainable (XAI) signals, and a signal lifecycle kept separate from execution: an execution intent is not an order.

A modular monolith was chosen over microservices for strong consistency and simpler deployment, while leaving a clean path to extract high-load components later. External providers — market data, brokers — sit behind adapters, and domain rules are testable without infrastructure.