Today's AI models are remarkably capable, yet they still invent figures, give different answers to the same question and tend to tell users what they want to hear. In most settings these are minor flaws. In investing, where decisions involve real capital, they are serious ones.

Vinley was therefore never designed as a wrapper chatbot connected to unreliable financial data. Over the past three years we have built a framework around Vinley's AI agent that keeps its answers grounded in verified data, consistent from one request to the next and open to scrutiny, while leaving every investment decision with the user.

We believe that investment grade AI carries the following attributes:

• Explainability answers the question of 'Why this answer?'. Every conclusion in Vinley shows the data, sources and calculation steps behind it, so the reasoning can be checked.

• Trustworthiness answers the question 'Can I believe it?'. Answers are grounded in verified financial data and filings. Zero fabricated data points.

• Reliability shall confirm 'Is it accurate?'. Numbers stay identical to the reference analysis (0% variance). Text stays within ±5%.

• Repeatability. The same question returns the same result, run after run, regardless of when or how often it is asked.
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   

How we make it safe
Language models handle language well but are unreliable at arithmetic. For that reason, none of the calculations in Vinley are performed by the model. Intrinsic value, the Beneish M-Score and every other metric are computed in our own engine, and the AI receives only the verified results, retrieved from data sources we curate and normalise ourselves.

Our curated knowledge base also takes precedence over anything the model learned during pretraining. In fact we do NOT allow Vinley's AI agent to rely on pretraining data that is not part of its own curated knowledge that we have provided and continuously maintain. The methods themselves are deliberately conservative and transparent. Terminal values are excluded from valuations, and results are accompanied by a sensitivity analysis rather than a single point estimate.

 
We clarify before we answer
A request such as "analyse Tencent" contains several unstated assumptions: the type of analysis required, which of the company's numerous listings is meant, and the date the analysis should reflect. Many AI tools simply choose on the user's behalf. Vinley retrieves the possible options in advance and asks the user to confirm them, so that the answer reflects the original intent.

For a specific ambiguous prompt like 'analyse Tencent' after having requested the user to confirm the ticker to analyse (disambiguation of Tencent), Vinley has been configured that such a prompt will trigger an analysis of the company fundamentals (scope disambiguation) and as of today ( time disambiguation).






























 
Four guardrails, built in

• No sycophancy. The same standards apply to every user, regardless of the answer they may be hoping for. Vinley will report that a company "fails 2 of 5 checks" rather than call it a great pick.

• No anthropomorphism. Vinley is a tool, not a persona. It does not express feelings or personal opinions, and it grounds its statements in evidence such as the company's annual filings.

• Plain language. Explanations are written so that any investor can follow them, with technical terminology kept optional. A valuation multiple, for instance, can be expressed simply as "$18 paid for every $1 of profit".
  Information, not recommendations. Vinley does not issue buy or sell calls. When asked whether to buy a stock, it presents the evidence and leaves the decision to the user.

These guardrails make a visible difference. In its default configuration, a general-purpose model told "I feel sad today" may respond with words of comfort and references to song lyrics. Vinley instead explains that it is designed to help users understand investments rather than to act as a companion, and redirects the conversation to the financial data.



























 
Repeatability
An investment analysis is only useful if it can be reproduced. General-purpose AI models increasingly rely on live internet searches to answer financial questions, so their results depend on which sources they happen to find at that moment. As a result, they often return different figures, or unusually wide valuation ranges, when the same question is asked twice.

In our own tests, the ranges they suggested changed from one run to the next even though the underlying data had not. Vinley avoids this by running every calculation in its own engine on curated data, so the same request produces identical numerical results each time, with only minor variation in wording (within ±5%). Investors can therefore revisit an analysis, track it over time or share it with others, knowing the figures will not change simply because the question was asked again.


























 

Examples of repeatability in Vinley
To show what repeatability means in practice, we asked Vinley the same question ten times: a fundamental analysis of Tencent Holdings (0700.HK). Below are runs 1, 5 and 10.
 
In the visual below, you can observe that the figures do not change. Each run opens with the date of the data, so it is always clear which point in time the analysis reflects. Next comes a company overview covering industry, country, currency and moat How is this possible ? Because we have trained Vinley to do so.

In all three runs, Vinley returns exactly the same figures for Tencent. Price to earnings is 14.17, price to free cash flow is 18.56, dividend yield is 1.23% and the Altman Z-Score is 9.87. These numbers are not generated by the language model. They are calculated in Vinley’s own engine and passed to the AI through retrieval from our curated data sources (RAG via API and MCP). Only the wording of the commentary varies slightly. A P/E of 14.17 is described as “moderate” in one run and “favourable” in another, but the underlying value is identical. An investor who returns to the analysis a week later will find the same results.

































In the second visual about repeatability below, you can observe the structure consistency in Vinley's answers because repeatability also applies to how an answer is organised. The assessment follows the same order every time: reliability of financials, relative valuation, passive income and returns to shareholders, then solvency and financial strength. This consistent layout makes the reasoning easy to follow and to check.

































Visuals, not just chat
Air traffic controllers do not manage flights through a chat window; they rely on displays that present the full situation at a glance. Research on human-AI interaction also suggests that heavy reliance on chatbots can create a sense of speed while reducing genuine understanding. Vinley therefore combines its AI agent with visual dashboards, allowing users to examine the evidence directly. More information about this can be found in this whitepaper. 

 
Regulary tested, publicly shared
Each quarter, Vinley is benchmarked against leading general-purpose AI models on real investment scenarios covering US and European equities, technical charting and international markets. The full report is published at vinley.ai/research, as we believe trust in financial AI has to be earned transparently.

 
What this means for you
Vinley's AI takes care of data gathering and calculations, discloses the source of every figure and reports the evidence objectively, including when the findings are unfavourable. It does not present itself as a companion or as a financial adviser. Judgement and decision-making remain with the investor.
Vinley provides information for educational purposes and does not give financial advice.