Daloopa

Extraction is the product, and it only works if it runs continuously. Models read filings, footnotes, investor presentations and transcripts and turn them into structured, source linked datapoints. The defining use case is keeping a model current within hours of a company reporting. Without the models, the accumulated dataset survives as an archive that decays immediately and the product stops existing, because earnings season is when the tool earns its place.

A buy side analyst founded the company and describes the problem as manually sourcing and entering data, which takes time away from actually analyzing investments. Displacing that manual work is the entire premise, not a feature on top of one.

Autonomy exposure is structurally low because the product supplies data rather than making decisions, and the analyst who receives it does the judging. What is undescribed is the process inside extraction. No confidence score is exposed on an extracted figure, no review queue or human verification step is described for ambiguous or non standard disclosures, and nothing states what happens when a company changes how it reports a metric mid history, which is exactly the case where automated extraction silently produces a discontinuity. The source hyperlink lets a user check any figure, but checking is the user's responsibility and no proactive flag on low confidence extraction is described.

Verification is a property of every individual output, not of the process around it. Each datapoint is hyperlinked to the filing, footnote, presentation or transcript it came from. Any figure a model produced can be checked in one click against an authoritative document the vendor did not create, and that applies universally rather than in principle.

A published benchmark reports agents reaching approximately 90 percent accuracy on this data against roughly 19 to 20 percent on public web sources. That is a quantified comparison in a category where accuracy is usually asserted. Independent commentary adds two qualifications. The benchmark does not specify which of three tested agent frameworks produced the full gain. The claim of ten times the data density of competing providers is vendor reported, without independent verification.

More than 160 institutional customers span hedge funds, mutual funds and bulge bracket banks. Revenue is stated to have more than doubled year on year, roughly $47 million has been raised, and a major investment bank is on the register as an investor. Coverage is quantified and checkable at more than 5,500 public companies with thirteen years of history. The most distinctive customers are not financial institutions at all.

Three leading AI platforms buy the dataset as grounding for their own financial products, a form of validation from buyers who evaluate data quality as their core competence. A published benchmark accompanies it, with agents reported to reach approximately 90 percent accuracy on this structured data against roughly 19 to 20 percent on public web sources.

Subject matter settles the cross customer training question. The dataset is public disclosure rather than customer data, so nothing belonging to one institution could improve a model serving a competitor. The safety architecture goes further. Every datapoint carries a hyperlink to the document it was extracted from, which is grounding by construction rather than by instruction.

The company positions the product explicitly against web sourced data that produces hallucinations and outputs that cannot be tied back to a trusted source. Nothing states how customer usage or query data is treated, and that is the one stewardship surface that does exist.

The privacy position is clean by structure, earned by subject matter rather than by policy. The entire dataset is built from public company disclosures: filings, footnotes, investor presentations and earnings transcripts. There is no personal data about individuals in the payload and no consumer whose records could be exposed. The open question is on the customer side.

What an analyst queries and models reveals their research direction, which is commercially sensitive and potentially close to material non public information about that firm's intentions. No statement covers how usage data is handled.

The dataset itself is built from public documents, so the consequence of a breach is lower than for vendors holding consumer or transaction data. Institutional buyers of this profile run vendor security reviews regardless. The newer business of supplying AI platforms also raises the question of what customer and query data those integrations expose. No attestation, certification, trust center or enumerated framework is published, so nothing lets a buyer assess it.

No supervisor, statute or instrument is named, which is partly appropriate. A data provider drawing exclusively on public disclosure carries no licensing obligation of its own and needs none. What is absent is engagement with the regimes the output enters. Sell side equity research is a supervised activity with its own record keeping and substantiation requirements. Data feeding a published research note, or a valuation used in a transaction, sits inside those obligations. Nothing published addresses how the platform supports them beyond the general claim that traceability meets compliance requirements without extra effort.

With no consumer subjects, protected class analysis does not apply. The governance question that replaces it is coverage. A universe of 5,500 companies necessarily excludes most listed issuers globally, and extraction quality will not be uniform across it. Companies with dense, well structured, English language disclosure are easier to read than smaller, non United States or less conventionally reported ones.

That means the issuers already least followed by analysts are the likeliest to be thinly or inaccurately represented in a dataset increasingly used to ground automated research. Nothing published describes which companies are covered, how the universe is selected, or how extraction accuracy varies across it.

Error correction is built into the product rather than promised alongside it. No guarantee or indemnity on data accuracy is published, so nothing commercially binds the vendor. Every datapoint links to the document it was extracted from, so a user can verify any figure against an authoritative original the vendor did not produce. That makes mistakes findable rather than latent, and it holds for every value. The limit is that finding an error is not the same as having it fixed. No correction, restatement or notification process is described for a datapoint discovered to be wrong, and nothing addresses what happens to models built on it downstream.

The data chain is fully disclosed by being fully public. Filings, footnotes, investor presentations and earnings transcripts are named as the sources, and anyone can obtain them, so no licensed feed, aggregator or proprietary supplier is in the path. On the model side, the protocol layer is described as agnostic and names the frontier platforms it interoperates with. Those are downstream consumers rather than upstream dependencies, and no provider is named for the extraction models themselves. No subprocessor list or hosting arrangement is published.

Machine addressable distribution is extensive. An MCP server exposes the dataset to language models and agents. It is explicitly model agnostic and integrated with a named financial services product from one frontier laboratory, and a custom assistant is published in another's public directory. Customers are stated to use the data directly inside three major assistant platforms and inside another vendor's research product.

Alongside that sits conventional analyst integration that delivers data straight into existing financial models. The same asset reaches both a human building a spreadsheet and an agent answering a question.

Residency carries less exposure than usual, because the underlying data is public disclosure rather than customer records, so what crosses a border is not sensitive in the usual sense. No hosting provider, region selection, residency commitment or private deployment option is published. Nothing says where customer query and model data rests, which matters for a global institutional client base with its own data governance requirements, and for European users in particular.

No pricing, packaging or basis of charge is published and every route in is a demo request. The question has real shape for this product because there are two very different buyers, an analyst team licensing seats against a company universe and an artificial intelligence platform licensing the dataset for grounding, and nothing indicates whether charging runs per seat, per company covered, per datapoint or per query. Category norm for institutional data, where the incumbents are equally opaque.

The buyer set covers both sides of the market: buy side hedge funds and mutual funds, and sell side equity research at bulge bracket banks. Private equity and valuation teams use the same history for comparable company analysis, and corporates use it for their own benchmarking. A newer segment sits outside financial services entirely, selling structured data to AI platforms that need verifiable grounding rather than open web retrieval.

The coverage is deep rather than broad across institution types. This is a capital markets research product, and it does not reach banking operations, lending, insurance or wealth advisory, which is a deliberate focus rather than a gap.