VERIFICATION INFRASTRUCTURE FOR REGULATED AI
Prove what
your AI did.
AI now writes, decides and advises inside regulated firms. When a regulator asks what happened, most firms have nothing to show. Velkron is the audit layer: deterministic rule checks now, cryptographic proof of inference next.
SELECTED · LIVE REPLAY
Product preview with sample data. The live console is at console.velkron.xyz.
Deterministic checks
Verdicts come from rules and logic, not a model's opinion. Same input, same verdict, every time, with the binding rules cited.
Learn more›Sealed receipts
Every run is hashed and chained. Months later, in an audit, you can prove exactly what ran and what it found.
Learn more›Proof of inference
Zero-knowledge proofs that a specific model made a specific decision, without exposing weights or data. In research.
Learn more›The AI stack has no audit layer.
From the EU AI Act to the FCA and the Gulf regulators, firms must show why AI output was compliant. A chat log is not evidence.
Billions went into making models produce more. Almost nothing went into checking what they produce.
In regulated work, an answer nobody can verify is a liability. The firms that deploy AI fastest will be the ones that can prove what it did.
We didn't invent the demand. Regulators on three continents wrote it down.
“A firm must ensure that a communication or a financial promotion is fair, clear and not misleading.”
“High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system.”
“An investment adviser may not include any untrue statement of a material fact in an advertisement.”
Every one of these is a live obligation. Velkron is the infrastructure that turns them from legal text into checks, records and proofs.
Four steps. No black box.
01
Submit
A promotion, a decision, a document. Dashboard or API.
02
Check
Deterministic run against an encoded rulebook. No opinions.
03
Verdict
PASS, FAIL with the breach named, or REFUSED to a human.
04
Receipt
The run sealed in a hash chain. Audit-proof, months later.
3
VERDICTS · PASS / FAIL / REFUSED
0
GUESSES · REFUSAL IS A FEATURE
SHA-256
EVERY RUN SEALED AND CHAINED
4
RULEBOOKS · FCA LIVE, GULF + EU NEXT
One call. Verdict, citations, receipt.
Drop verification into an existing approval flow, a CMS or an agent pipeline. If your AI publishes, the call goes in front of it.
REQUEST
POST /v1/checks
Authorization: Bearer vk_live_...
{
"artefact_url": "https://firm.com/promo.pdf",
"rulebook": "uk-fca",
"modules": ["COBS4", "CONC3"],
"mode": "pre_publication"
}RESPONSE · 200
{
"verdict": "fail",
"breach": {
"rule": "COBS 4.12A.34R",
"reason": "incentive to invest"
},
"citations": ["COBS 4.12A.34R"],
"grounding": "sufficient",
"receipt": {
"sha256": "c81f30aa…77c2",
"chained": true
}
}IDEMPOTENT · DETERMINISTIC · SAME INPUT, SAME VERDICT, SAME RECEIPT
One verification stack. Wedge first.
Others add AI to compliance. Velkron adds proof to AI, starting where the regulatory pain is sharpest. The same layer then extends to every regulated vertical.
Checks financial promotions against a regulator's rulebook before publication. Returns pass or fail with the exact rules cited, plus an evidence pack a compliance officer can file. When the rules don't support a verdict it refuses. It never guesses. First rulebook live: UK FCA. The engine is rulebook-agnostic; ADGM and DFSA are next.
A tamper-evident record of every AI decision: model, input, output, sealed in a hash chain. Edit the history and the chain breaks in plain sight.
Zero-knowledge proof that a specific model produced a specific decision, without exposing the model weights or the customer data.
One engine. Every regulated market.
A rulebook is content, not code. The verification engine stays the same; each jurisdiction is an encoded rule module with its own citations. Statuses below are honest: one live today, the rest in the order the market is pulling us.
EUROPE
United Kingdom
FCA
European Union
AI ACT · ESMA
Switzerland
FINMA
MIDDLE EAST
Abu Dhabi · ADGM
FSRA
Dubai · DIFC
DFSA
Saudi Arabia
CMA
AMERICAS
United States
SEC · FINRA
Canada
CSA
Brazil
CVM
ASIA-PACIFIC
Singapore
MAS
Hong Kong
SFC
Australia
ASIC
India
SEBI
LIVE = IN PRODUCTION · NEXT = IN BUILD ORDER · HORIZON = ROADMAP · THE WAITLIST SETS THE ORDER →
THE WAITLIST SETS THE BUILD ORDER
Your jurisdiction, sooner.
Rulebooks are built in order of demand. Join with your work email and jurisdiction: every signup moves that rulebook up the queue, and you get first access when it goes live.
THE POSITION
Trust is an assertion.
Proof is a receipt.
Every AI vendor asks regulated firms for trust. Velkron is built so nobody has to give it. Check the output, cite the rule, seal the record. That is the whole company.
The toolkit is public.
Four small Python libraries, zero dependencies, one idea: check the output, don't trust it.
Calibrated uncertainty. Models that know when to abstain.
Checks LLM output against the sources it claims to cite.
Evolutionary search where only verified candidates survive.
Multi-agent consensus with trust that persists across runs.
Satyawan Singh
ML ENGINEER · SOLO FOUNDER
Every entry in this build log was shipped by one person in the last 18 months. The pattern is the point: every project makes machine output provable.
4 open-source verification libraries
abstention, grounding, verified search, agent consensus
119M-parameter speech model
trained from scratch for UK clinical audio
MSc dissertation, University of Leicester
trust scoring for clinical AI
140-crate Rust blockchain
25,000+ passing tests, consensus verification primitives
RegTech retrieval engine
provenance-gated citations for FCA-regulated firms
Asked and answered.
Q1Is this another AI compliance copilot?+
No. Copilots generate suggestions with a model and ask you to trust them. Velkron's verdicts come from deterministic checks against an encoded rulebook. Where a model assists internally, it only proposes; the deterministic layer decides, and every verdict carries its rule citations.
Q2What happens when the system is not sure?+
It says so. Below a grounding threshold, Velkron returns REFUSED and escalates the case to a human, with its reasoning attached. The refusal rate is reported to you, not hidden. We consider this the most important feature in the product.
Q3Which rulebooks are supported today?+
UK FCA financial promotions (COBS 4, CONC 3, including the COBS 4.12A crypto regime) is live. ADGM FSRA and DIFC DFSA are next on the roadmap, followed by EU AI Act conformity evidence. The engine is rulebook-agnostic; a new jurisdiction is content work, not a rebuild.
Q4What does a pilot involve?+
You send a set of recent promotions, we return full evidence packs, and your compliance team judges the output against their own review. Pilot terms are agreed per firm. If the packs are not useful, you will know within a week.
Q5Is our data used to train models?+
No. Checks are deterministic and your artefacts are processed for verification only. Receipts are hashes; they do not contain your content and cannot be reversed into it.
Q6Who is behind Velkron?+
A solo technical founder with a public build record in AI verification: open-source libraries, a clinical speech model, a formally tested blockchain and a RegTech retrieval engine. The full log is on the company page.
Taking on a small group of design partners in regulated finance.
Wherever you're regulated: if your firm publishes financial promotions, or is putting AI anywhere a regulator can see, write to us. We reply to every request within two working days.