What it does

Every three months, big companies like PepsiCo hold an earnings call. Their executives tell investors how the business is doing and what they expect next. One sentence can move the stock price.

On the call, executives often talk about a change without repeating the number. The number is in the company’s press release, so listeners must stop and look it up. On October 8, 2026, PepsiCo’s finance chief said “it’s the margin piece that’s driving the guide down in EPS”. In plain words: “We now expect profits to grow less than we said before.” The exact number was in that morning’s press release, not in the sentence.

Call Desk listens to the call live. About 12 seconds after that sentence, a red card pops up: PepsiCo lowered its profit forecast. About 12 seconds later, the card fills in:

• Before and now: profit per share was expected to grow about 5% this year. Now it is 2.5% to 3.5%. The new number is labelled “not spoken”, because it comes from PepsiCo’s press release that morning, not from the call.
• Proof: the press release, found live on the web, and the forecast PepsiCo gave three months earlier.
• The stock: how PepsiCo’s stock moved, next to similar food and drink companies.
• How sure: a confidence score of 98%. The evidence gave 92%, then two agents checked again and both agreed.

Think of a sports announcer who shouts “that changes the score!” but never says the score. Call Desk is the friend beside you who checks the scoreboard and says: “It was 5, now it’s 3. Here’s a photo of the scoreboard.”

It is not built for one company. On Delta Air Lines’ call, the finance chief reads the forecast out loud. Call Desk shows that Delta now expects to earn $5.10 to $5.60 per share this year, down from a range of $6.50 to $7.50. It also shows that Delta lowered its forecast of cash left over after running the business (free cash flow), from a range of $3–4 billion to about $2.5 billion.

When the call ends, Call Desk writes a short summary that you can paste into an email.

How we built it

Five tools work together. Each one has one job.

1. Listen (Gemini). The call plays at real-time speed in 12-second pieces. Google’s Gemini 3.8 Flash writes down each piece, labels who is speaking, and picks out claims. It gives each claim an importance score from 0 to 1: how much it could move the stock. Claims scoring 0.6 or more become cards, one card per topic. A faster streaming listener on the Gemini Live API is also built. It cuts the time to a card to about 5 seconds, and it is off by default.

2. Check (Gemini and three tools). An AI agent built on Gemini calls three tools at the same time:

   • Snowflake remembers what the company forecast last time (Cortex Search).
   • Nimble searches the live web for the company’s press release and news coverage.
   • Tiger Data measures how the stock moved, with one SQL query.

   A second Gemini call then writes the card’s answer.

3. Score. Confidence = 0.45 × old-forecast check + 0.35 × web check + 0.20 × stock check. A failed tool lowers the score. It never stops the call. After the answer, Snowflake’s agent and Nimble’s agent check again. Each one that agrees adds 3 points, each one that disagrees takes away 10, and the score never goes above 98%. In the demo, both agents agree, so the profit card goes from 92% to 98%.

4. Keep a record (Sentry). Sentry records every agent run, AI model call and tool call. Each card links to its record.

5. Show. A Python web server (FastAPI) streams everything to a one-page app. Every run is recorded, so a replay mode can play it back with the original timing.

Results (live, all five tools on)

• PepsiCo answer key: 8 of 8. We wrote the answer key from the company’s own press release.
  • Three checks cover what was said on the call: the lowered profit forecast, pressure on profit margins, and next year’s forecast being put off until February.
  • Five checks cover the numbers. Each forecast line must have the right value and the right direction (up, down or unchanged). The lines are two measures of profit per share, two measures of sales, and the tax rate.
• Delta answer key: 10 of 10. It catches the lowered full-year forecasts for profit per share and for cash left over after running the business (free cash flow). It also gets the forecasts for the last three months of the year (profit per share, profit margin and fuel cost) and Delta’s year-end debt level.
• Speed: the red card appears about 12 seconds after the finance chief’s sentence, and the full answer about 12 seconds after that (the run we show on stage, 20261011-150448). With the Gemini Live listener, the card appears about 5 seconds after the sentence.
• Replay recordings: no pipeline errors, and no cards where the checking agent failed. Some optional Nimble specialists (for example the news search) were unavailable at times. The card shows them as unavailable, and the main checks still ran.

Challenges

• PepsiCo’s call was questions and answers only, so nobody read the forecast numbers out loud. We taught the agent to find them in that morning’s press release and to label them “not spoken”.
• Cutting the audio into 12-second pieces once split “$5.10 to | $5.60” in half. Each piece now starts 2.5 seconds early, and a later mention completes a number that was cut off.
• The memory of earlier forecasts must not know anything from after the call, or the check would be cheating. An early file let one later number slip in, and we removed it.

What’s next

• Watch several companies’ calls side by side.
• Send alerts to Slack or Teams.
• Explain what each change means for the person reading it.
• Make the faster streaming listener the default once it labels speakers as well as the 12-second listener does.

AI tools used

Claude Code (Claude Opus 5.5) wrote and tested the code, researched the sponsor SDKs and prepared the demo data, under the team’s direction. whisper.cpp located the demo clips offline. Seer, Sentry’s AI debugger, explained the cause of one bug after we had fixed it. It read only the public repository and wrote no code. Inside the product: Gemini, plus the Snowflake Cortex models listed in the README’s “Tools used” section. The demo video’s narration voice was generated with Gemini TTS.