ABI NexaIntelligence OS
The data & access layer
ΣFundamentals & Filings DataStandardized and as-reported, citedfxExcel Add-inThe cited number in your model{ }Data APIThe whole corpus, programmatic∞MCP ServerThe AI-native connector⛱Research WorkspaceThe whole file, per name
AI-native deliverables
ΣOperating & Valuation ModelsBuilt, sourced, refreshed on the print▤Company Tearsheets & ProfilesThe as-of-date one-pager✎Confidential Information MemorandaThe sell-side document, drafted✎Investment MemorandaThe buy-side write-up, traceable▣Pitch BooksProfiles and precedents, quoted⚟Screening & Idea GenerationScreens across the corpus
Who we serve
BBuy-SideHedge funds & asset managersSSell-Side ResearchEquity researchIBInvestment Banking & Corp DevM&A and coveragePEPrivate Equity & VCDeal teams, diligence to model▣CorporatesStrategy, IR and FP&A{ }Data & Quant TeamsAgents and pipelines
Plans Methodology Resources Sign in Request access
Products
Fundamentals & Filings DataStandardized and as-reported, citedExcel Add-inThe cited number in your modelData APIThe whole corpus, programmaticMCP ServerThe AI-native connectorResearch WorkspaceThe whole file, per name
Solutions
Operating & Valuation ModelsBuilt, sourced, refreshed on the printCompany Tearsheets & ProfilesThe as-of-date one-pagerConfidential Information MemorandaThe sell-side document, draftedInvestment MemorandaThe buy-side write-up, traceablePitch BooksProfiles and precedents, quotedScreening & Idea GenerationScreens across the corpus
Clients
Buy-SideHedge funds & asset managersSell-Side ResearchEquity researchInvestment Banking & Corp DevM&A and coveragePrivate Equity & VCDeal teams, diligence to modelCorporatesStrategy, IR and FP&AData & Quant TeamsAgents and pipelines
PlansMethodologyResources
Sign inRequest access
Methodology

Everyone asserts accuracy. We publish ours.

We ran the audit, we publish the method, and the harness reports its own exceptions, which is the part that makes the rest believable. Here is exactly what we measure and how.

Request accessSee the platform
100%

Zero wrong numbers

479 figures re-checked against the regulator's own structured record. Every one matched, at a 95% confidence interval of 99.2 to 100%.

99.8%

Cited to the right filing

The number is right, and the document we attribute it to is the one the regulator records it in.

100.0%

Identity integrity

Every fact's key recomputes from its own content, with 0 mismatches across 29.8M facts and no sampling.

99.88%

Balance integrity

Assets equal liabilities plus equity. The exceptions are surfaced and listed, not hidden.

Structural integrity is measured on the full corpus and is exact. Source-verified accuracy is measured on a fleet-wide sample against the regulator's structured records (SEC XBRL in the US, and the NSE record for India-filed names) and reported with its sample size and confidence interval. The audit harness and the sample are both re-runnable, and the rate is refreshed as coverage grows.
The data contract

Six rules we do not break.

Enforced in the product, not promised in a brochure. Each rule exists because the alternative quietly produces a plausible, wrong answer.

1A gap is never a zeron/d, never blank, never 02Native currency, keptFiled currency and scale, crore stays crore3Each market, its own formUS-GAAP, Ind-AS, bank Schedule III4Vintages never mixedRestated and as-first-reported, switchable5Derived figures say soAnything computed carries a marker6Coverage is statedA public page, market by market, gaps includedThe datacontractsix rules we keep
When a model answers

The model never does the math.

The same rules hold when the answer comes from an agent. The model retrieves and explains; it does not invent the numbers, and it cannot pass off a figure it did not fetch.

The math runs in the engine, not the model

Valuations and calculations run in a deterministic engine that reproduces the same result byte for byte. The model narrates that result and checks it for sense; it never produces the number itself.

Every figure resolves to a source

A guardrail checks that each number in an answer traces back to a retrieved cell in the corpus. If it does not resolve, the answer is refused rather than guessed.

Grounded, with the filing attached

Retrieval is scoped to the companies, periods and documents you asked for, and every claim carries the filing, accession and location it came from.

Tested before it ships

A fixed set of questions with known answers runs on every model and prompt change, so a regression is caught by us, not by you.

Trust it because you can check it.

Every figure on every surface names the filing behind it. Request access and click one.

Request accessSign in