Veyra

AI financial intelligence that verifies evidence before producing company research, risks, and scenarios.

Comment

Inspiration

Financial decisions affect everyone, but high-quality financial research is still difficult to access and even harder to trust.

Today, a person can ask an AI model, “Should I worry about this company?” and receive a fluent answer. But in finance, fluency is not enough. Users need to know:

  • Where did the facts come from?
  • Was the company correctly identified?
  • Is the data current?
  • What evidence is missing?
  • Were the calculations checked?
  • What could make the thesis wrong?

Veyra was built from that idea: financial AI should not just answer; it should prove how it reached the answer.

The inspiration was to create a financial intelligence system that feels useful for an 18-year-old learning investing, a 70-year-old trying to understand a company, a founder analyzing a private business, a CFO reviewing risk, or an analyst pressure-testing a thesis.

What Veyra Does

Veyra is a ChatGPT-native financial intelligence system built on MCP.

Instead of acting like a generic chatbot, Veyra works like an evidence-gated financial analyst. It resolves the company, gathers available evidence, checks what data is missing, runs structured analysis, challenges the thesis, and returns a traceable answer.

Veyra can:

  • Resolve public and private company identities
  • Search current public company evidence
  • Analyze SEC Companyfacts and regulatory data
  • Analyze private companies from user-supplied evidence
  • Produce company risks, scenarios, thesis changes, and evidence gaps
  • Run deterministic financial calculations
  • Explain provider failures instead of hiding them
  • Generate replayable decision traces
  • Support paper-trading simulation without real trade execution
  • Separate public-safe tools from the full developer-mode system

Veyra is research-only. It does not provide personalized investment advice and does not execute real trades.

Why This Matters

Most people do not need another financial dashboard. They need a system that can explain financial reality in a way they can trust.

A retail user may ask:

“Should I worry about Apple?”

An analyst may ask:

“What changed in the last 30 days, and does it affect the thesis?”

A founder may ask:

“Analyze my private company from this pitch deck and financial summary.”

A CFO may ask:

“What risks, dependencies, and scenarios should I monitor?”

Veyra is designed to serve all of these users with the same principle:

No unsupported claim should look more certain than the evidence allows.

How It Works

Veyra follows an evidence-first architecture:

  1. Resolve the entity Veyra first determines what company or security the user means.

  2. Build a capability manifest It shows what analysis is possible: identity, filings, market data, news, valuation, risks, scenarios, portfolio impact, and execution status.

  3. Gather evidence It routes to available providers such as SEC data, current public web evidence, configured feeds, and user-supplied documents.

  4. Create verified claims It separates supported facts from unsupported assumptions.

  5. Run deterministic analysis Financial ratios, valuation readiness, scenario analysis, risk checks, and paper simulations are handled by code, not guesswork.

  6. Challenge the thesis Veyra generates counterarguments, null hypotheses, and evidence gaps.

  7. Apply quality gates The output is checked for provenance, freshness, math validity, safety, and missing-data handling.

  8. Return a trace The user can see what happened, what failed, and what evidence supported the conclusion.

The guiding formula is:

$$ \text{Trustworthy Financial AI} = \text{Evidence} + \text{Provenance} + \text{Verification} + \text{Honest Uncertainty} $$

What Makes It Different

A normal AI answer can sound confident even when the evidence is weak.

Veyra is different because it is designed to say:

  • “This is verified.”
  • “This is missing.”
  • “This provider failed.”
  • “This company is ambiguous.”
  • “This module is not applicable.”
  • “This conclusion should not be made from the available evidence.”

That is especially important in finance, where a wrong answer can create real harm.

Veyra does not silently convert missing financial values into zero. It does not invent company identifiers. It does not use synthetic fixtures as live evidence. It does not pretend a private company has a public ticker. It does not place real trades.

How We Built It

Veyra was built as a TypeScript and Node.js MCP server connected to ChatGPT Developer Mode and deployed on Render with Docker.

The system includes:

  • MCP server endpoints: /mcp and /mcp-public
  • ChatGPT Developer Mode connector
  • Global company resolver
  • Capability manifest engine
  • Evidence provider routing
  • SEC Companyfacts support
  • Optional OpenAI current web evidence
  • Private company intelligence mode
  • Current company investigation mode
  • Deterministic financial analysis modules
  • Quality gates
  • Replayable traces
  • Paper trading simulation
  • Disabled real-execution boundary
  • GitHub deployment workflow
  • Render cloud deployment

The architecture was built around one core idea: the AI model should not be the whole financial system. The model should sit inside a controlled system of evidence, schemas, deterministic calculations, validation, and safety gates.

Challenges We Faced

The hardest challenge was not generating financial text. The hardest challenge was preventing the system from sounding certain when it should not.

We had to handle:

  • Missing prices
  • Missing revenue
  • Missing cash and debt
  • Unknown tickers
  • Ambiguous company names
  • Private companies without public filings
  • Provider outages
  • API quota failures
  • Public market data licensing limits
  • Live data freshness
  • Synthetic fixture boundaries
  • Paper trading versus real execution

For example, if a price is unavailable, Veyra must not calculate a fake order notional. If a company cannot be resolved, Veyra must not create a fake ticker. If live market providers are unavailable, Veyra must not invent a catalyst.

That reliability work became one of the most important parts of the project.

What We Learned

We learned that financial AI needs more than intelligence. It needs discipline.

A useful financial AI system needs:

  • Identity resolution
  • Evidence provenance
  • Freshness tracking
  • Deterministic calculations
  • Quality gates
  • Replayability
  • Clear limitations
  • Provider status reporting
  • Strong safety boundaries

We also learned that “I cannot verify this yet” can be a powerful product feature. In finance, honest uncertainty is better than confident fiction.

Who It Helps

Veyra can help:

  • Everyday users understand companies in plain language
  • New investors learn how to think about risk and evidence
  • Financial analysts pressure-test investment theses
  • CFOs and operators monitor risks and scenarios
  • Founders analyze private company evidence
  • Reviewers audit how a conclusion was produced
  • Institutions build toward more transparent AI-assisted research workflows

The long-term vision is not just another stock assistant. It is a financial reasoning layer that people can trust because it shows its work.

What’s Next

The current version is a working hackathon prototype. The next production milestones are:

  • Persistent PostgreSQL storage
  • OAuth and user accounts
  • Tenant isolation
  • More global regulator adapters
  • Broader company identity coverage
  • Licensed market and news providers
  • Portfolio syncing
  • Production privacy and terms
  • OpenAI public app review
  • More polished ChatGPT widgets
  • Institutional monitoring and audit workflows

Current Status

Veyra is deployed, connected through ChatGPT MCP, and ready for demo and review.

It is not yet a public-production financial platform, but it proves the core idea:

AI financial research can be evidence-gated, traceable, honest about uncertainty, and useful to both everyday users and serious analysts.

Built With

Share this project:

Updates

Submission history