Case Studies/VaultIQ+
AI Product Engineering · End to End

From days of due diligence to a cited risk report in minutes

VaultIQ+ is an AI-powered AML and compliance intelligence platform we designed, built, and operate for InsightX. It turns the slow, manual work of vetting a person into a structured report across eleven risk categories: every claim traceable to its source.

VaultIQ+Built for InsightXRegTech / compliance intelligence
Days → Minutes
Research time per subject
11
Risk categories per report
100%
Findings cited to source
vaultiq.app
Report ready
Overall risk signalMedium
Litigation history
Medium
Watchlists & sanctions
Clear
Adverse media
Elevated
Regulatory issues
Clear
Source 1 Source 2
11
Specialised AI agents
6+
Licensed data sources
Minutes
From search to report
Full
Audit trail on every report

Overview

A complete due-diligence platform, not a chatbot

A compliance officer enters a person of interest: name, country, date of birth, optional images and identifiers. Within minutes, VaultIQ+ returns a structured risk report that aggregates findings across eleven distinct categories, from litigation history and sanctions exposure to adverse media and political connections.

Each section is summarised by AI from primary sources, and every summary links back to the original evidence an analyst can verify. The human stays in control of the decision; the AI does the heavy reading. It is a real multi-tenant SaaS (identity, role-based access, billing, role-scoped dashboards, document-grade PDF export) built to the standard a regulated enterprise buyer expects.

Who it serves

  • Compliance & AML teams
  • KYC and due-diligence analysts
  • Risk & onboarding functions
  • Anyone vetting a person before they transact

What we owned, end to end

Product strategy & UXMulti-tenant SaaS backendApplied AI & agent engineeringCloud architecture & data engineeringIdentity, security & DevOps

The Challenge

The bottleneck was never the information. It was reading it

Across compliance and due diligence, a specialist has to learn everything publicly knowable about a person, fast, defensibly, and in a format another professional will trust. In practice that meant toggling between ten data sources, copy-pasting findings into a document, and hoping nothing slipped through.

Days per file

A single thorough background report could take a senior analyst one to several days of manual research.

Inconsistent quality

Depth and presentation varied by analyst and by how much time the deadline allowed.

The “did we miss something” risk

With evidence scattered across separate subscriptions, no one could be certain a material finding wasn’t overlooked.

Senior time on junior work

Expensive specialists spent their best hours assembling documents instead of making judgement calls.

How It Works

Three coordinated layers, running in parallel

When a search is submitted, VaultIQ+ fans out across three layers on managed, event-driven cloud infrastructure, so no analyst waits on a single long-running script, and a slow data source can never block the whole report.

01

Data acquisition

Multiple licensed compliance sources are queried at once: court and litigation records, sanctions and watchlists, news archives, government records, KYC data, each in its own pipeline so one degraded source can’t stall the rest.

02

Knowledge indexing

Returned documents are converted into a private, per-search knowledge index, so each agent can ask its sources the right question instead of re-reading everything from scratch.

03

Reasoning & synthesis

A fleet of eleven specialised AI agents (one per risk category) retrieves the most relevant evidence and produces a structured finding: a summary, a risk signal, and explicit citations back to the source.

Search submitted
Data acquisition
Knowledge index
11 agents
Report assembled
Delivered & notified

Event-driven and orchestrated in the cloud: every step independently observable, retryable, and parallel.

The Agent Fleet

Eleven specialised agents, not one giant prompt

One monolithic prompt is a smell. VaultIQ+ decomposes the problem into eleven agents, each with a narrow job, its own retrieval scope, output schema, and evaluation criteria, so any category can be improved or debugged without touching the others.

Red flags

Surfaces the highest-priority risk signals across all evidence.

Litigation history

Civil and criminal cases from court and archive records.

Watchlists & sanctions

KYC/AML hits against sanctions and watchlist databases.

Regulatory issues

Enforcement actions and regulatory findings.

Adverse media

News, scandals, and reputation signals from archives.

Political connections

Positions, donations, and political affiliations.

Corporate connections

Company affiliations and business relationships.

Ethics & conduct

Conduct concerns and ethics-related findings.

Integrity

Integrity signals weighed across the evidence pool.

Social media presence

Public social footprint across platforms.

Sponsored & paid content

Sponsored and paid-content footprint.

A final assembly step composes the eleven findings (plus an AI-written executive summary) into the report the analyst reads, then notifies the right people the moment it’s ready.

Continuous Improvement

The system gets sharper, under the client’s control

A capable AI product isn’t “ship and forget.” VaultIQ+ includes an AI Tutoring layer that lets InsightX’s own experts steer report quality over time, without writing code, and without any change reaching production unmeasured.

Experts correct the AI, not engineers

Super-admins capture structured observations against any report section, propose a revised instruction, and roll it out. The whole tutoring loop is operable in-app, with no engineering involvement.

Every change is measured before it ships

No new prompt version activates without being scored against a ground-truth corpus for accuracy, format compliance, and consistency. The quality score is shown at the moment of activation, and any version can be rolled back.

In the client’s voice from day one

InsightX’s own example reports are baked in as guided examples, so output matches house structure, tone, and editorial conventions from the first report, not after months of correction.

Inspectable, multi-stage reasoning

Each agent works in visible steps: select, analyse, validate, format, with automatic retry on failure, and the full chain is recoverable for any report: which evidence, which prompt version, which examples, and what the model returned.

vaultiq.app / ai-tutoring
Topics
Red flags
Litigation
Sanctions
Adverse media
Regulatory
Litigation · prompt v4Pending review
Quality vs. ground-truth corpus
94%
Accuracy
100%
Format
97%
Consistency
Approve & activate
Roll back

The tutoring loop

Observe
Triage
Draft revision
Evaluate vs. corpus
Approve
Activate / roll back

How We Build AI

The patterns that make AI safe to deploy in a regulated domain

VaultIQ+ is one product in one domain, but the discipline behind it ports to any “knowledge-worker bottleneck” problem.

LLMs summarise, never the source of truth

We retrieve evidence with deterministic queries and use the model to summarise, classify, and structure it. Retrieval-augmented generation as a discipline, not a buzzword.

Decompose, don’t monolith

Specialised agents with narrow jobs and individual evaluation. One agent regressing never poisons the rest.

Orchestrate in the cloud, not in a request

Long-running AI pipelines belong in a real workflow orchestrator: every step independently observable, retryable, and parallel.

The audit trail is the product

“The AI told me” is not a defensible answer. Every summary cites the primary document behind it.

Humans in the loop where stakes demand it

Routine screening is automated; higher-stakes subjects route through an approval-gated full-report workflow.

Evaluate, don’t just ship

Known inputs, expected behaviours, regression flagging. You can’t unit-test an LLM, but you can catch regressions before users do.

Outcomes

What changed for InsightX

Framed in structural terms we stand behind: the platform’s design guarantees, not invented percentages.

Days → Minutes
Research turnaround

A report that took a senior analyst one to several days now arrives in minutes.

11 categories
Consistent every time

The same structured coverage across every subject, regardless of analyst or deadline.

100% cited
Verifiable findings

Every AI summary links back to the primary evidence, defensible to a regulator.

Analyst → decision-maker
Higher-value work

Specialists stop assembling documents and start making judgement calls.

Client Feedback

At InsightX, we have been very fortunate to work with ObjectSingle Technologies and I cannot recommend them highly enough. They developed, from scratch, a secure platform with associated CMS for our reports and we are delighted with it. Always professional, responsive and friendly.

Justin Williams

Co-CEO (Product), InsightX

ObjectSingle has worked with us on the development of an editorial management platform and client interface. From initial brainstorming to the delivery of the final product, they worked as an integral part of our team. Always proactive, responsive, and available for support.

Veronica Ferrari

Head of Insight, InsightX

Beyond VaultIQ+

The same architecture solves a whole class of problems

Wherever a knowledge worker is paid to read everything publicly knowable and produce a defensible recommendation, the VaultIQ+ pattern applies.

KYB & vendor onboarding

Multi-source data fusion and AI summarisation, applied to corporate entities instead of people.

M&A & investment research

An agent fleet over filings, news, patents, and ESG sources, theses instead of risk categories.

Insurance underwriting

Agents that read submissions, policy documents, and prior claims into a risk assessment the underwriter can act on.

Legal discovery & contract review

Section-level summaries with citation back to the exact clause.

Internal knowledge platforms

“Ask the AI, get cited evidence” over your own document repositories.

Common Questions

What buyers ask us about building AI like this

We never let the model be the source of truth. Evidence is retrieved deterministically from real data sources; the AI only summarises and structures it, and every claim cites the document it came from. In a compliance setting, an uncited “finding” simply doesn’t ship.

Have a knowledge-worker bottleneck worth automating?

If your specialists spend their best hours reading documents instead of making decisions, the architecture behind VaultIQ+ likely applies. Let’s map it to your problem.

30-min call
NDA available
Response within 24 hours