Theia Insights

The removal test leaves the static taxonomy this product exists to replace. Proprietary models combining natural language processing, quantitative modelling and knowledge graph architecture read regulatory filings, earnings transcripts, press releases and financial statements to produce a continuously updated multidimensional description of a company, expressing it as percentage exposures across several themes simultaneously rather than as one label, with weightings adjusting automatically as new information emerges. That output cannot be produced by rules, and the founding team came from artificial intelligence research at a major technology company rather than from financial data.

The platform supplies structure rather than conclusions, and the classification it produces flows directly into portfolio construction, index membership and trading decisions at the institutions using it, with weightings updating automatically as new information arrives rather than on a review cycle. Auditability is claimed as a property of the output.

What is absent is any description of oversight over the updating itself: nothing states whether a material reclassification is reviewed before publication, whether customers are notified when a company's exposures shift significantly, or what happens when the model and a human analyst disagree about what a business does.

Auditability is central to the positioning rather than incidental, with the layer described as enabling consistent, auditable and economically grounded decisions, and grounding classification in observable corporate disclosure means an output can in principle be traced to the filing that produced it.

The company also frames its purpose as a model risk argument about other people's systems, that artificial intelligence deployed in research and trading requires accurate economic inputs and without them risks producing confident but structurally incorrect outputs, which is an unusually precise statement of the failure mode it sells against. What is missing is measurement of its own: no agreement rate against expert classification, no stability analysis on how often weightings move, and no validation result is published.

Adoption is described by institution type rather than by name, with the technology stated to be in use by global index providers, major banks, large asset managers and hedge funds for research, portfolio construction, risk measurement and trading, which is a demanding buyer set for a company founded in 2022.

One commercial relationship is named: an early channel partnership with a major exchange group, which is distribution through an institution whose own business depends on classification being right. An eight million dollar Series A closed in March 2026 led by an early stage fintech fund, bringing total funding to 14.5 million. What holds this at B is that no individual customer is identified and the exchange partnership is characterised as early.

The pooling question is largely closed by construction, since the corpus is public disclosure that belongs to nobody and no customer supplies proprietary data to generate the classification, so one institution's use cannot advantage another through the training set.

Commonality is in fact the stated objective rather than a risk, with the company positioning itself as building a shared ontology for financial markets and its investor noting that shared definitions unlock network effects, which means every customer receiving the same view is the product working as intended. Held at B because nothing states whether customer usage, queries or feedback inform the models, and in a classification business what customers ask about is itself informative.

Structurally favourable by subject matter, which is the Daloopa and Nammu21 position. Every stated input is public corporate disclosure, covering regulatory filings, earnings transcripts, press releases and financial statements, so there is no personal data in the payload and no customer records are ingested to produce the classification. That removes most of what this axis exists to interrogate.

Held at B because no data processing terms, retention schedule or subprocessor list was located, and the private markets expansion will introduce material that is not public and may carry confidentiality obligations.

No attestation, certification, trust centre or enumerated framework was located. Index providers, major banks and large asset managers all run vendor assessment programmes as a condition of onboarding, and a channel partnership with an exchange group implies a further layer of review, so the assurance exists privately and none of it is published for a prospective buyer to read.

No supervisor, statute or instrument is named. The omission is pointed because of who the customers are: index providers construct benchmarks, and the inputs, methodology and governance of benchmarks are directly regulated in several of the markets served, with requirements covering methodology documentation, change control and oversight. A classification layer feeding index construction sits inside that framework, and nothing published addresses it or any data licensing position for the filings and transcripts ingested.

No individual is assessed and the adapted exposure is consequential, because classification determines capital allocation. A company placed into a favoured theme attracts index inclusion, thematic fund flows and analyst coverage, while one classified out of it loses them, and here those weightings update automatically as new information emerges, which means a company's sector exposure and therefore its passive flows can shift without any human decision, notice or right of reply.

A second exposure follows from the inputs: models reading filings, transcripts and press releases will describe a large company with extensive English language disclosure far more accurately than a smaller or non English reporting one, so classification quality tracks disclosure resources. No coverage breakdown or accuracy analysis across company size, geography or reporting language was located.

No guarantee, indemnity or falsifiable commitment was located. The institutional customer is reasonably served, since classification grounded in public filings can be checked against those filings and the platform's auditability claim supports that.

The company being classified has nothing at all, and its position is materially consequential rather than merely uncomfortable: a business assigned exposures that misrepresent what it does may be excluded from index inclusion and thematic flows, which affects its cost of capital, and no notification, review or correction route is described anywhere.

The input side is disclosed with unusual clarity and every source named is public: regulatory filings, earnings transcripts, press releases and financial statements, which lets a buyer understand exactly what the classification is derived from and verify any output against the same documents. The technical architecture is also described rather than obscured, covering natural language processing, quantitative modelling and knowledge graph structure. What is not disclosed is the model layer itself, with no provider named for any component, and no subprocessor list or hosting arrangement located.

The layer is designed for consumption rather than as a destination, with four tools stated to integrate into existing investment workflows, data systems and artificial intelligence driven analytical platforms, which is the right architecture for infrastructure meant to sit underneath other software. The named exchange channel partnership provides distribution through an institution that already supplies data to the same buyers.

What is not published is the mechanics: no interface documentation, data delivery format, coverage universe size or named platform integration was located, so an institution cannot assess what adoption involves.

No hosting provider, region selection, residency commitment or private deployment option was located. Exposure is lower than for platforms holding client data, since the corpus is public disclosure, and it is not absent, because customers consuming classification data reveal through their queries what they are researching, and institutions in several served markets impose their own requirements on where such activity is processed.

No pricing, packaging or basis of charge was located. Four distinct tools are described as forming the intelligence layer, and consumption ranges from an index provider licensing classification data at scale to a fintech platform embedding it, which are unlikely to be priced the same way. Nothing indicates whether charge falls per company covered, per seat, on data volume or by licence.

Six buyer types are addressed, spanning index providers, asset managers, hedge funds, banks, fintech platforms and, with the new funding, private equity firms and exchange traded fund issuers. Functional reach covers research, portfolio construction, risk measurement and trading, so the same layer supports several parts of an investment process rather than one.

Coverage to date is public companies globally, with expansion into private markets stated as the use of the round and justified on the basis that no dynamic classification exists there at all. The limit is that this is a data layer rather than an application, so its reach depends on what customers build on it.