Vesuliria AVT · Analyze · Verify · Trace

AI analytics with evidence you can show an auditor.

Ask questions of your institution's data in plain language. Vesuliria AVT returns the answer together with where every number came from, which statistical test it passed, and a signature. Months later, the same result can be recomputed. It runs inside your environment; data never leaves.

Request a 15-minute demo Pilot program
0unsourced (invented) numbers — structurally blocked
1.00 / 0.40true-driver detection (F1): Vesuliria / naive method
2 / 100false alarms when there is no signal (target ≤ 10)
On-premiseyour servers, your model; data stays inside

The problem

“Chat with your data” tools are fast. But indefensible.

In typical AI analytics tools the model writes numbers into the text itself, redefines metrics every time and mistakes the largest difference for a cause. The answer looks right, but when an auditor asks “where did this number come from?” there is nothing to show.

Typical AI + SQL tool

  • The model writes numbers itself; invented figures are possible
  • Redefines “non-performing loan ratio” on every question
  • Presents noise as findings and correlation as cause
  • The same question gives different answers on different days
  • Once data changes, old answers cannot be verified

Vesuliria AVT

  • The model cannot write digits; every number comes from a recorded query
  • Metric definitions are the ones your institution approved
  • Every claim passes a statistical test; otherwise it says “inconclusive”
  • Same question, same data: bit-identical numbers
  • Reports are signed and can be recomputed months later

How it works

The AI reasons. Deterministic code computes.

The model understands the question, forms hypotheses and writes the text. Numbers are produced by tested, audited code. This separation makes every figure traceable.

Understands

Resolves which metric and period you mean. If ambiguous, it asks instead of guessing.

Checks the data

Missing periods, broken records or drift stop the analysis or lower its confidence.

Tests

Pre-registers hypotheses; difference-in-differences, power analysis and multiple-testing correction eliminate false findings.

Writes with evidence

Every number is linked to its source; the report is signed and logged.

Auditability

Months later, “where did this number come from?” still has an answer.

SIGNATURE

Signed reports

Every report is digitally signed and sealed. A single changed character fails verification.

RECOMPUTE

Same data, same code

A copy of the source data and the calculation code can be archived; “Recompute statistics” reproduces the result.

TRACE

Evidence chain

From finding to test, query and SQL in one click. Rejected hypotheses stay on record with their reasons.

Enterprise

Runs inside your institution, under your rules.

Data never leaves

The AI model can run on your own servers. Outbound calls can be disabled entirely by configuration.

Enterprise identity

SSO (OIDC) sign-in and SCIM provisioning of users and groups from Entra ID / Okta.

Role-based access and sharing

Analyst, auditor and admin roles. A single finding can be shared without exposing the full report or queries.

Tamper-evident audit trail

Every view, share and deletion is written to a hash-chained log. EU AI Act mapping included.

Measured results

Measurements, not claims.

Measured on a synthetic credit portfolio with known true drivers and on public real-world datasets.

MeasureResult
True-driver detection (F1) — naive “largest difference” method: 0.401.00
Invented (unsourced) numbers0
Correct behavior: answer / clarify / refuse (111 questions)111 / 111
False alarms with no signal (100 trials)2 / 100
Core question set with a small on-premise model (7B)27 / 27
“No driver” case on real credit-card data (UCI)Correctly identified

Measurements are on synthetic and public datasets; validation on institutional data is the purpose of the pilot program.

Who it is for

For teams that must defend their numbers.

Risk and credit analytics

Tested answers to “why did non-performing loans rise, and which segment drove it?”

Internal audit and control

Source, test and signature for every figure; independent recomputation later.

Reporting and management

Removes the “the AI made it up” risk from analyses going to regulators and the board.

Banks, participation banks, finance, factoring and leasing companies, insurers and regulated fintechs.

Pilot program

We are looking for our first design partners.

The product works and has been measured; now we want to validate its value on real institutional data together. The pilot runs inside your institution, under your control.

6–8 week pilot
  • One dataset, 5–10 real business questions
  • On-premise installation; data stays inside
  • Metric definitions agreed with your team
  • A measured evaluation report at the end
  • Special terms for early partners

Contact

Let us show you in 15 minutes.

A live demo with your own example question, and an honest assessment of whether a pilot makes sense.

unlu.mehmett1@gmail.com LinkedIn