ACTIVE / LOCAL
Interactive prototype
Review the seven-screen synthetic workflow and download a local evidence manifest.
Launch interactive prototypeEntity-side evidence integrity / human accountability
Show the evidence behind a reported field. Stop when the record disagrees.
KRISEVA ATTEST is an entity-side evidence-integrity and human-accountability layer for AI-assisted regulatory reporting in GIFT IFSC. The demonstration compares candidate values, exposes conflict or missing support, and records a named human decision before anything is treated as settled.
Built for compliance teams at fund management entities (FMEs) in GIFT IFSC assembling evidence for periodic IFSCA returns. Fields are illustrative of those reported in FME periodic returns under the IFSCA (Fund Management) Regulations, 2025.
Launch interactive prototypeExhibit A / disagreement retained
One field. Two sources. No invented winner.
SOURCE A · PAGE 2
USD 25,000,000 Administrator statementREPORTED FIELD
CONFLICTING Human reason requiredSOURCE B · CELL C11
USD 24,000,000 Subscription register extractBoth candidates remain visible. The deterministic trace proposes no compliance judgment; a reviewer accepts, corrects, rejects, or defers with a reason.
Demonstration method
Compare / Escalate / Record
-
01 / COMPARE
Keep source and value together
Three governed fields resolve to supported, conflicting, or unsupported evidence states.
-
02 / ESCALATE
Stop at conflict or absence
Unequal values remain conflicting. Missing evidence produces an explicit abstention. The same evidence record powers compliance assembly, deterministic risk indicators, and fraud-analysis flags, each closed by a named human.
-
03 / RECORD
Attribute the human decision
The receipt retains candidates, source fingerprints, the reviewer, the reason, and unresolved items.
Submission index
Six artifacts, one controlled front door.
ACTIVE / LOCAL
Wireframes
Inspect the information architecture across desktop and mobile review states.
View wireframesVERIFIED / LOCAL
Demo video
Watch the captioned 90-second walkthrough of the bounded synthetic workflow.
Watch verified demoVERIFIED / LOCAL
Pitch deck
Open the rendered ten-slide presentation; its editable source remains in the package.
View verified pitch deckACTIVE / LOCAL
Technical notes
Read the static architecture, deterministic state flow, and operating limits.
Read technical notesPUBLISHED / PUBLIC
GitHub repository
Browse the source, tests, synthetic fixture, and reproducible build instructions.
View source repositoryOperating boundary
What this demonstration is allowed to show.
Synthetic and deterministic
The case is fictional. The disclosed trace makes no live model call.
No filing
The prototype does not write or submit an official return.
No regulator connection
It is not connected to IFSCA systems or DRR interfaces.
Human decision required
Conflict remains unresolved until a named reviewer records a reason.
Pilot agentic step
Model-driven source extraction and candidate proposal, operating inside the same PROPOSE / ABSTAIN / DECIDE boundary shown by the deterministic trace.
Operating scope / no integration claimed
Where ATTEST sits, and where it stops.
SOURCE
Underlying books, registers, statements
KRISEVA ATTEST
Evidence acquisition, extraction proposals, conflicts, human decision, manifest
REGULATORY SUBMISSION CHANNEL / DRR
Taxonomy, submission, validation, supervisory consumption
Entity-side, before submission. No DRR connection claimed.
Build responsibility
Founder-owned engineering and integration.
The repository, deterministic review logic, interface integration, verification, and publication gates remain founder-owned. No additional public team role is asserted on this surface.
Anticipated challenge / answered in advance
Hard questions, straight answers.
What is actually different here?
Not a feature list: an accountability contract. Public materials already show adjacent capabilities, lineage, audit trails, exception workflows, and agent evidence logs, at vendors including Regnology, Workiva, IRIS, Duco, and Hadrius; we do not claim otherwise. The differentiated object is the combination: exact source region, every conflicting candidate preserved, a named human reason before a conflict closes, and a portable manifest, applied to FME regulatory reporting. That full combination is not located in the public materials reviewed. We call it a hypothesis, not a validated moat.
Why can't ChatGPT or Copilot do this?
A generic assistant answers a prompt once. It does not persist a source binding, retain conflicting candidates, force a named decision before export, or emit a tool-independent manifest. That is a workflow-state design, not a model-quality gap.
Does DRR make this redundant?
DRR is submission-side infrastructure; ATTEST sits upstream, proving which source supported a field before it reaches a DRR form. We claim no gap: equivalent source-region provenance from DRR or an administrator is a pre-registered kill condition for this wedge.
Where is the agentic AI?
Nowhere in the running code today, and the screens say so. What exists is the contract any agent must obey: propose a value with exact references, abstain when evidence is missing, and let only a named human decide. IFSCA's July 2026 AI Survey names source-referenced document extraction as an observed IFSC AI use case and human-in-the-loop oversight as the most widely adopted production safeguard; that is the shape of the model we plan to put inside this boundary.
Is human-in-the-loop an admission the AI failed?
No. In regulated reporting the accountable human is the requirement, not a fallback: someone signs under their own name. The design makes that decision explicit, attributed, reasoned, and separately confirmed under maker-checker review.
Why trust this team?
KRISEVA already ships evidence-first, human-review procurement software in Indian defense today. That process discipline transfers: source binding, human sign-off, an auditable trail, under a demanding audit culture. Fintech domain knowledge is what the residency provides; the claim discipline here is machine-enforced.