Ask in plain language and five data agents go out at once — filings, earnings, market data, news, macro. Three analysis passes work over what comes back, a corroboration check flags whatever the sources disagree on, and every claim in the report links to where it came from.
One query in. Five sources at once, three kinds of analysis over what they return, and a corroboration check before any of it reaches the report.
An analyst types what they want to know — a company, a sector, a thesis to test. An orchestrator reads the intent and decides which sources the question actually needs, so a question about this week's coverage doesn't drag in a decade of filings and a macro pull. Fewer sources, faster answer, same rigour where it matters.
Claude API · orchestration and selective dispatchFive specialised agents run at the same time rather than in a queue: filings from EDGAR, quarterly EPS history with beats and misses and the next reporting date, price history and valuation multiples, recent news with sentiment scored, and the macro series that frame the whole picture. This is the step that turns an afternoon into a few minutes — the work was never hard, it was just serial.
SEC EDGAR · yfinance · FRED · news and sentimentThe same collected data gets read three different ways. A quantitative pass computes ratios, trends, and anomalies. A qualitative pass interprets management tone and strategic direction — the part of an earnings call that never shows up in a number. A risk pass pulls the risk factors out of the filings and news and sorts them into categories.
Claude API · quant · qualitative · risk extractionFindings are checked against each other across sources before anything is written up. Where the sources agree, confidence is high and the claim stands. Where they contradict each other or only one thin source supports a conclusion, it gets flagged for the analyst rather than quietly averaged into a confident sentence.
confidence scoring · contradictions surfaced, not smoothedA structured research report where every claim links back to the filing, figure, or headline it came from. Each run is persisted, so the same query re-run next week becomes a comparison rather than a fresh start — and the report lands in Slack where the desk already is.
Postgres run history · Slack delivery · REST accessResearch nobody can check is research nobody can act on — and a plausible, fluent, unsourced paragraph is worse than no paragraph, because it reads like work that was done. So citation isn't a formatting nicety here, and neither is the corroboration gate: together they're what lets an analyst put their own name on the output.
Each agent owns one source and knows its shape. The orchestrator activates only the ones a given question needs.
Annual and quarterly reports plus material event disclosures. The primary record — slow to arrive, but the thing everything else is ultimately checked against.
EPS history, beats and misses against expectations, and when the next print lands. The surprise pattern often says more than the absolute numbers.
Price history and valuation multiples, which is what turns a good business into a question about whether it's a good position at this price.
The fastest-moving input and the least reliable one — which is exactly why it's corroborated against filings rather than trusted on its own.
The backdrop a single-name thesis lives inside. Cheap to include, expensive to have ignored when the whole sector was moving for one reason.
One orchestrator, three layers of agents underneath it, and a gate before anything is written up.
Interprets the query and decides which agents run. Everything below it is fan-out and fan-in around this one decision.
Five agents, one source each, running simultaneously. One slow source delays the run; it doesn't block the other four.
Three passes over what came back, each answering a different kind of question about the same underlying evidence.
Compares findings against each other and scores confidence. Disagreement becomes a flag rather than an average.
Every claim carries its citation into a structured report. Runs are persisted, which is what makes the same query repeatable as a weekly tracker rather than a one-off.
Six modular blocks. Sources and delivery swap per client; the fan-out and the gate don't.
Reads the query and picks the sources.
Five feeds pulled at the same time.
Quant, qualitative, and risk over one dataset.
Checks the sources against each other.
Binds every claim to where it came from.
Run history kept, report sent where the desk is.