Equity research,
settled by AI debate.
A committee of Bull, Bear and Quant agents builds the case on any listed company, then argues it out in a rebuttal round — with your own research thrown into the debate. A Consensus agent weighs what survives into two views: a short-term outlook anchored on the next earnings print, and a 12–36 month viability thesis. Run one name or a whole portfolio, and keep questioning the reasoning until you trust it.
For portfolio managers · research analysts · stock pickers · hedge funds
Inputs
Bull agentBuilds the strongest honest case for owning the name — demand signals, pricing power, the company's own beat history, and the catalysts management is most likely to lean on into the next print.
Bear agentArgues the other side — margin pressure, channel build, share loss, and the guidance risk a consensus estimate tends to under-price. The case a short seller would bring to the meeting.
Quant agentReads the market's own forecast: implied move from the at-the-money straddle, max-pain strike, skew, and the historical distribution of surprises.
Your researchDrop in your own notes, channel checks or house view. It enters as a first-class input the agents are required to engage with — not a footnote.
Debate
DebateEvery thesis lands in one shared context. Overlapping claims collapse, contradictions surface, and each argument stays attached to the evidence behind it.
Rebuttal
RebuttalBull and Bear read each other's case and answer it directly. Claims that cannot survive the challenge get dropped; the ones that do carry more weight downstream.
Consensus
ConsensusWeighs the surviving arguments by evidence strength and agent conviction, then answers on two horizons: where the next print is likely to land, and whether the business holds up over years. You can keep questioning it afterwards.
Outputs
Short-term outlookThe near-term view, anchored on the next earnings print: which way the committee leans — beat, meet or miss — the confidence it puts on that, and what the options market is already pricing. This is the half that gets scored against the actual report, whichever way it lands.
Long-term viabilityThe other half of the answer: a research thesis that outlives the quarter. Business viability and competitive moat, long-horizon bull and bear cases, the catalysts and risks behind them, every claim carrying an evidence weight — and versioned, so you can diff it against the last one.
Follow-up
Ask the agentNeither view is the end of it. Question the assumptions, ask which evidence moved the number, push back on a factor you think is wrong — the consensus agent answers from the same context it decided on, and keeps answering.
What the agents read before they argue
Run on the shared defaults, or connect your own data sources — drop in your NewsAPI, Alpha Vantage or EarningsAPI keys, and your own Anthropic, OpenAI or Google model key, and every run uses yours. Keys are encrypted at rest and never shown back to you in plaintext.
Five steps, no black box
You can follow the reasoning at every stage — which agent said what, what evidence it leaned on, where the consensus disagreed with it, and what it says when you push back.
Pick a company, or bring the whole book
Search any listed name, or upload a holdings file and send positions into a single debate or a batch run. The calendar knows when each one reports, so you can run it days ahead of the print or the morning of.
Agents pull their own evidence
Fundamentals, filings, headlines and the live options chain — fetched at run time, not from a stale snapshot, and cited back in the output. Connect your own provider keys and they use those instead.
They argue, rebut, and converge
Three cases in parallel, then a cross-examination round, then a consensus agent that has to justify the weight it gave each side.
You question the result
Two views land: a short-term outlook on the coming print and a 12–36 month viability thesis. Neither is final — challenge an assumption, ask which evidence moved the number, and the consensus agent answers from the context it decided on.
The earnings call gets scored
Every short-term call is written down before the report lands and graded against it afterwards with a Brier score. Wrong calls stay on the record.
Four agents, and one seat that's yours
Each one is prompted to be good at its own argument, not to be agreeable. The disagreement is the point — a call that survives it is worth more than one nobody challenged.
Bull agent
BUYSIDE CASEBuilds the strongest honest case for owning the name — demand signals, pricing power, the company's own beat history, and the catalysts management is most likely to lean on into the next print.
Bear agent
SHORT CASEArgues the other side — margin pressure, channel build, share loss, and the guidance risk a consensus estimate tends to under-price. The case a short seller would bring to the meeting.
Quant agent
OPTIONS · MAX PAINReads the market's own forecast: implied move from the at-the-money straddle, max-pain strike, skew, and the historical distribution of surprises.
Consensus
CONFIDENCE-WEIGHTEDWeighs the surviving arguments by evidence strength and agent conviction, then answers on two horizons: where the next print is likely to land, and whether the business holds up over years. You can keep questioning it afterwards.
Your research
OPTIONAL HUMAN INPUTDrop in your own notes, channel checks or house view. It enters as a first-class input the agents are required to engage with — not a footnote.
Graded on the record, not on the pitch
A research process is only worth what its scoreboard says. The earnings call is the part of each analysis that can be checked within weeks, so every one is timestamped before the print and graded after it — including the ones that go badly.
Written down first
The direction and the confidence are persisted before the company reports. Nothing is edited after the fact.
Brier score, lower is better
A call graded on confidence, not just direction: (confidence − outcome)². Being loudly wrong costs more than being unsure.
Scored daily, automatically
A scheduled job reconciles open predictions against actual reported results every morning and publishes the result.
Questions worth asking first
Who is it for?
Anyone whose job is forming a view on a stock: portfolio managers, research analysts, stock pickers, hedge fund teams, and serious individual investors. Use it on one name ahead of a print, across a whole uploaded portfolio in a batch run, or as a second opinion on a long-term position you already hold.
How is Talakay different from other products?
Most AI research tools put one model in front of a data feed and hand back a summary. Here, agents with opposing mandates analyse the same aggregated evidence separately — fundamentals, filings, news and the options chain — and then have to answer each other, with your own research entered as a case they must engage with rather than a prompt they can ignore. The Consensus agent consolidates what survives into a short-term and a long-term view, and keeps the full debate as context, so you can question why it weighed things the way it did instead of taking the answer on trust. The earnings calls are recorded before the report and scored afterwards, so there is a track record to check rather than a claim. And it runs on your book as well as on single names: upload a portfolio and send positions straight into analysis.
Is it only about earnings?
No. Earnings are the anchor for the short-term outlook, because a reported quarter is the one thing a near-term view can be graded against. The same debate also feeds a 12–36 month thesis on business viability, competitive moat, catalysts and risks, which is about whether the company is worth owning rather than how one print lands.
Is this investment advice?
No. It is a research tool that shows you an argued case and the agents' own confidence in it. Nothing here is a recommendation to buy or sell anything, and every prediction can be — and sometimes is — wrong.
Why a debate instead of one model?
A single prompt tends to launder its own assumptions: it finds evidence for whatever it decided first. Forcing a bull case, a bear case and a market-implied case to be written separately, then to answer each other, surfaces the disagreement instead of averaging it away.
Where does the data come from — and can I connect my own?
Market and fundamental data from Yahoo Finance, filings and XBRL company facts from SEC EDGAR, headlines and sentiment from news providers, and live options analytics — implied move, max pain, skew — from the options chain. Sources are fetched at run time and cited in the output. You can also connect your own: NewsAPI, Alpha Vantage and EarningsAPI keys go in your settings, are encrypted at rest, and take priority over the shared defaults for every run on your account.
How do the short-term outlook and the long-term thesis differ?
They answer different questions off the same debate. The short-term outlook is a single-quarter lean on the coming print — beat, meet or miss — scored once the company reports. The long-term thesis is a 12–36 month structural view — business viability, moat, long-horizon bull and bear cases, catalysts, risks, each claim carrying an evidence weight — and it is versioned, so you can compare it against the last one and see what changed.
Is the first answer the end of it?
No. Both views open into a conversation with the consensus agent: challenge an assumption, ask which piece of evidence moved the confidence number, or argue a factor was weighted wrong. It answers from the same context it decided on, so the reasoning stays inspectable rather than being re-improvised.
What does the confidence number actually mean?
It is the consensus agent's stated probability that its direction is right, and it is scored as such. A 70% call that lands wrong is penalised more than a 55% one, which is the whole point of grading with a Brier score rather than a raw hit rate.
Can I add my own research?
Yes. Your notes go in as an input the agents have to engage with directly, and the rebuttal round will push back on them the same way it pushes back on the Bull and Bear cases.
Which model runs the agents?
Configurable per account — Anthropic, OpenAI or Google. You can supply your own API key, which is encrypted at rest and never returned in plaintext.
Put your next idea in front of the committee.
Pick a company or upload a portfolio, watch four agents fight over it, and get a short-term outlook and a long-term thesis you can interrogate line by line.
One company, debated, every Monday
Each week we pick a company reporting that week, run the full debate with rebuttals, and email you the thesis and the report. Unsubscribe in one click.
AI-generated research, not investment advice. See our privacy policy.