CR-04
2026 · Q2
Entity resolution with stated reasoning
Match scores hide their logic. We study resolution systems that emit human-readable justifications as a first-class output — and measure whether analysts trust them correctly.
RegSentia — AI-native compliance for financial institutions
The AIfor compliance.
RegSentia screens, investigates, and drafts the filing — and shows its work. Sealed, reviewable evidence behind every decision. Delivered as SaaS or deployed inside your environment.
6 weeks
median deployment
0 bytes
exported from your perimeter
100%
decisions with sealed evidence
Anywhere
SaaS · VPC · on-prem · air-gapped
01 — Mission
Financial trust runs on infrastructure no one has rebuilt in twenty years. We are rebuilding it.
02The problem
Compliance software is broken in ways everyone in the industry accepts as normal.
F-01ACCEPTED FAILURE
Implementations measured in years
Legacy compliance platforms are sold as software and delivered as consulting projects. Eighteen-month rollouts are normal. Some never finish.
F-02ACCEPTED FAILURE
Your data, copied into someone else's cloud
Every incumbent vendor requires institutions to export customer records — names, transactions, histories — into the vendor's environment. The most sensitive data a bank holds, duplicated by default.
F-03ACCEPTED FAILURE
Decisions no one can explain
A match score of 87 is not an explanation. When a regulator asks why an account was cleared, “the system said so” is the answer most institutions actually have.
F-04ACCEPTED FAILURE
Evidence assembled by hand, after the fact
Audit trails are reconstructed from screenshots and spreadsheets, months later, under deadline. The record of a decision should be created at the moment of the decision.
None of this is a law of nature. It is the residue of architecture decisions made two decades ago — and never revisited.
03The approach
We build infrastructure, not dashboards.
Five components, designed together, deployable apart. Hover any part of the system — each one explains itself.
RegSentia platform — schematic
Customer perimeter — everything above runs inside it
CRN-SCComponent spec
Screening Core
Sanctions, PEP, and adverse media screening with reasoned matches. Not string similarity — entity resolution that states why two records are the same person, and why they are not.
IN: entities, watchlists · OUT: reasoned match decisions
04 — First principle · Zero-copy
01 / 03
Your data stays where it is.
Customer records, transactions, histories — under your keys, in one place. Nothing is exported to a shared vendor cloud. Ever.
02 / 03
The engine moves to the data.
RegSentia ships as a single artifact that installs inside your environment. The software travels. The data does not.
03 / 03
Evidence accumulates in place.
Screening runs, decisions are made, and every one writes sealed evidence — all without a single byte leaving your control.
P-02
Evidence-first
Every decision writes its own audit trail at the moment it happens. Inputs, reasoning, model versions, human overrides — sealed, append-only, queryable. An audit becomes a query, not a quarter.
P-03
AI-native
Models are load-bearing, not decorative. Rules written in plain language. Investigations drafted by agents. Reports composed automatically. And every conclusion shows its reasoning — because in compliance, an answer without reasoning is not an answer.
P-04
Runs anywhere
One codebase, one artifact: managed cloud, your VPC, your datacenter, air-gapped if it must be. Deployment is a property of the architecture, not a separate product.
P-05
Weeks, not years
Deployment is an install, not a consulting engagement. Sensible defaults, policy packs per jurisdiction, migration tooling that reads your existing rules. Live in weeks — then tuned forever after.
05Product
It starts with screening.
RegSentia Screening resolves entities against sanctions lists, PEP registers, and adverse media — and states its reasoning for every match and every clearance. Fewer false positives is a side effect. Explainability is the point. RegSentia does the work; your team decides, with the evidence in front of them.
Screening decision — sample outputentities are fictional
subject ANDRIN VELKOV · b. 1971 · account 4471…
list hit A. VELKOV — consolidated sanctions, entry 8812
verdict ESCALATE · confidence 0.94
Reasoning
- · name similarity 96.2% — transliteration variant, Cyrillic source
- · date of birth exact match
- · nationality consistent with list entry
- · counter-evidence checked: none found
- · evidence sealed — EVD-9107, reviewable in full
Trajectory — toward the operating system for regulated institutions
NOW
AML Screening
Sanctions, PEP, adverse media — reasoned matches with sealed evidence.
NEXT
Transaction Monitoring
Plain-language rules compiled to executable detection logic.
NEXT
Case Management
Investigations that assemble themselves around the evidence.
LATER
Regulatory Intelligence
Continuous mapping from regulation to your controls.
LATER
Compliance Agents
Agents that investigate, draft, and file — with humans deciding.
LATER
Risk Infrastructure
The substrate every regulated workflow runs on.
06 — The record
Four minutes from alert to filing. With the paper trail written as it happened.
ALERT04:12:07
- screening hit · subject 8842-K
- list: consolidated sanctions · entry 8812
- confidence 0.94 · auto-escalated
INVESTIGATION04:12:09
- agent: resolving entity graph…
- agent: 3 related accounts · 2 jurisdictions
- agent: counter-evidence search — none found
- draft summary ready for review
EVIDENCE04:13:41
- EVD-9107 · inputs, reasoning, model v4.2
- human review: confirmed · analyst M.R.
- record sealed · append-only
REPORT04:15:02
- narrative composed from sealed evidence
- format: regulator schema · jurisdiction A
- attachments compiled · 11 artifacts
FILED04:16:20
- submission package exported
- acknowledgment received
- elapsed: 4m 13s
07Research
The hard problems are research problems. We treat them that way.
CR-03
2026 · Q1
Compiling policy language to detection logic
Compliance rules are written in prose and enforced in code, with translation loss in between. We explore direct compilation from controlled policy language to executable, testable rules.
CR-02
2026 · Q1
Evidence structures for regulated AI decisions
What must a record contain for a model-assisted decision to be defensible years later? A proposed minimal evidence schema: inputs, versions, reasoning, and the human boundary.
Research notes published as the work matures. No launch-day whitepapers.
08Engineering
How we build.
E-01
Every decision explains itself
If a component cannot state why it did something, it does not ship.
E-02
Modules assume nothing
Every capability works alone. No module requires another to exist.
E-03
One artifact, every environment
If it cannot run air-gapped, the design is wrong — not the customer.
E-04
Replaceable by design
Customers can swap any part of the system. Including ours.
E-05
Boring where it counts
Novelty budget is spent on the product, not the plumbing.
E-06
Evidence is the schema
The audit trail is not a feature added later. It is the data model.
09 — Why RegSentia exists
Every institution we spoke to told us the same story. The compliance platform took months of heavy lifting to deploy — moving data, hand-configuring rules — and a year more before the team could trust what it produced. It holds a copy of their most sensitive data. Nobody can explain its decisions. And when the regulator calls, a team spends weeks reconstructing what the system should have recorded on its own.
The technology to fix this exists. The incumbents cannot use it — their architecture, their pricing, and their consulting revenue all depend on the old way. So we are building it properly, from the beginning.
— RegSentia, 2026
10Hiring
A small team, building carefully, for a very long time.
We hire people who find compliance interesting the way cryptographers find locks interesting — as a hard systems problem with real stakes. Remote-first. Async-heavy. Evidence over opinion, in the product and in the team.