The financial story humans need three reads to surface — extracted in one pass. Paste an earnings call transcript across any sector. One structured-output call returns the headline number grid, growth drivers with magnitude, headwinds with severity, capital allocation signals, a Q&A breakdown that grades how directly management answered each analyst, and the forward guidance against consensus. The same shape a junior IR analyst would build over an afternoon, generated as fast as you can pick a sector.
Browser
└─→ POST /api/lab/chat
- system: financial-extraction prompt with the schema
below (versioned: earnings.v1)
- user: the call transcript, verbatim
- temperature: 0.2 (low — extraction must be stable)
- max_tokens: 2800
← single response, parsed as JSON: {
company_profile, summary, quarter, tone,
key_numbers[], growth_drivers[], headwinds[],
capital_allocation[], strategic_priorities[],
qa_highlights[], guidance
}
← rendered as a number-card grid + drivers/headwinds
row + allocation/priorities row + a Q&A list with
directness grades + a guidance callout
One LLM call covering a finance-grade taxonomy. The model
extracts numbers verbatim, attributes them to a period and
a delta, and grades each Q&A on how directly management
answered — direct / hedged / deflected. The
guidance field is the most-watched field in
any earnings call; the model surfaces both the numbers and
their relation to consensus when stated.
Earnings calls are the densest information event in the public-markets calendar. A buy-side analyst might cover 30 names; their calls are stacked into the same five-day window each quarter. The first-pass synthesis — what changed, what's the guide, what did management dodge — is mechanical work that gets done at 11pm by tired humans.
This demo collapses that pass. Drop the transcript in (S&P, Bloomberg, FactSet all export plain text), get back the structure. The analyst still does the thinking — but starts the thinking from a populated workbook, not a blank one.
Same shape works for IR teams prepping the next call (what did our peers commit to that we haven't?) and corp-dev teams doing target diligence (what's the capital-allocation cadence over the last eight calls?). For investor relations specifically, the Q&A directness grade is the single most-requested feature — because the questions you dodge become next quarter's investor concerns.
Honest caveat: the model is an extractor, not a judge. Numbers can be mis-attributed across periods if speakers blur them. The "directness" grade is the model's interpretation, not a consensus opinion. Production deployments should overlay this with the firm's existing rule-based extractors and keep the human analyst in the loop for the high-stakes decisions.
Analyze a call to see telemetry.