OmniTrader

A trading product where configurable LLM-backed agents score market and social signals (Twitter and Telegram feeds over gRPC) before anything executes. The interesting part is not the trade; it is deciding, with an audit trail, whether to trade at all.

For whom: a Primus Cloud Solutions product. Constraints: every agent run must be attributable and costed, metrics must be first-class, and the Python agents and TypeScript services must share one contract.

• A Python agent service, a NestJS central service. Python where the models and signal libraries live; NestJS for auth, realtime eventing, cost accounting, and the REST and gRPC APIs.

• gRPC between them. One protobuf contract, generated on both sides.

• Prometheus metrics from day one. Agent latency, token spend, and signal counts are dashboards, not log greps.

• Cost accounting per agent run. Every model call is attributed to an agent, a strategy, and a user.

Still boarding: the agent service, control plane, and metrics are in place; execution is gated until the signal scoring proves itself on paper. Team of four.

Not in production long enough to have a proper incident. The first thing I would fix: agent cost accounting is computed from the model response after the fact, so a runaway strategy can spend a whole budget in one loop before the cap applies. Pre-flight budget reservation per run is the next change.