An analyst's day is dominated by reading: filings, transcripts, news, broker research, and increasingly alternative data. The first job of the system is not insight; it is triage. Which of the thousand documents that arrived this morning changes anything for the names this desk covers.

That is a retrieval and classification problem before it is a generation problem, and it is where most of the engineering effort goes. Entity resolution alone (which company, which subsidiary, which ticker) is a project in its own right.

We structure the system as narrow agents with explicit hand-offs: an ingestion agent that normalises documents and links them to entities, a monitoring agent that watches for defined event types, an analysis agent that drafts a claim with cited passages, and a review agent that checks the claim against the evidence before anything reaches a human.

Every hand-off is logged. An analyst, or a compliance officer six months later, can open a signal and see the documents, the passages, the model versions, and the intermediate outputs. The explanation is not generated after the fact; it is the trace.

The output of the system is a structured claim: what changed, for which entity, with what evidence, and how confident the system is given the quality of the sources. Confidence is calibrated against outcomes over time, and the calibration is visible to the desk.

Analysts rank the signals and can dismiss them with a reason. Those reasons go back into the evaluation set. The system learns what the desk considers noise, and the desk learns which sources the system is over-weighting.

Sensitive sources, internal research, and the desk's own positions cannot leave the institution. The retrieval layer and the models that touch that data run inside the network, and open-weights models serve well here. Public sources can be processed by hosted models where policy allows, with the split enforced by the ingestion layer rather than by convention.

Latency is less critical than in live sports, but freshness matters. A signal on an overnight filing that arrives after the open has missed its window.

Before go-live, compliance wants to know that the system cannot be steered by a single source, that the evidence trail is immutable, that model changes are controlled, and that analysts remain accountable for what they act on. Designing for those four questions from the start is faster than answering them under pressure at the end.

The measure of success is that analysts spend their mornings on analysis, and that the signals they act on can be explained to a client on one page.