Cortex Sentinel case study

Role: Product lead: research, specs, console UI and pitch
Company: Agentic AI Build Week 2026
Period: Hackathon · Agentic AI Build Week 2026
Stack: Go, React, PostgreSQL, Elasticsearch + yente

The problem

The insight

The bottleneck isn't detecting risk. It's changing the rules.
Cortex Sentinel pitch deck, p.3

GoTyme had just handed its banking customers a crypto wallet (6.5M users, the bank's own public figure). Every , withdrawal and cross-exchange transfer can raise an alert, so alerts grow with volume. A digital bank can't answer ten times the volume with ten times the analysts, and for a licensed crypto exchange one AML miss is a licence event, not a fine.

The tension sat in 2,000 alerts analysts had already resolved. Some alert types were cleared every single time: pure noise. One, high crypto-wallet risk, split roughly 50/50: a coin flip only a person should call. So the real question was never "can AI score this?" It was "which alerts are we allowed to automate, and can we defend that to a regulator on any date?"

Agentic AI Build Week 2026: 2nd place
resolved alerts in the team's corpus (seed data): 2,000
GoTyme users in the brief (company's public figure): 6.5M

Source: Cortex Sentinel pitch deck, p.2, p.3 and p.16. The user figure is GoTyme's own.

Ownership

What I owned

Thao owned

This was a team build on top of the open-source Marble decision engine. I led the product side: I researched Vietnamese AML law and the , wrote a 14-file user-story set, and specced every workflow as a clickable mockup before any UI existed. Then I re-skinned the console to match those mockups and wrote the pitch. Teammates built most of the backend and AI plumbing.

Two decisions were mine. First, spec before build: each screen existed as a mockup with real Vietnamese AML cases (fictional people) so the team argued about the workflow, not the pixels. Second, label every claim: in the pitch each number is tagged live, prototyped or , and the deck says out loud which parts were the platform's. Filter the map below to see exactly what that meant.

Thao owned

Disposition copilot mockup: a case with a 0.86 confidence score, a recommended escalation and cited reason codes — Two of my seven workflow mockups, specced before the console was built. All names are fictional.

STR drafting mockup: a queue of suspicious-transaction report drafts ranked by deadline — Two of my seven workflow mockups, specced before the console was built. All names are fictional.

Rule authoring

How a rule is born

Today a new laundering pattern means an engineer, a ticket and a blind deploy. In the console an analyst describes the rule in plain English, an agent turns it into a typed rule and checks it against the data model, and a person saves it. Step through it yourself.

Platform

Before it ships, the new version runs beside the live one

Demo data

71%

19%

Hover a segment, or tap a name below, to read it.

A candidate rule version is scored on the same traffic as the live version, with no effect on customers. This is how its decisions split in the demo run.Source: Cortex Sentinel pitch deck, p.7 (shadow test run, candidate v2 vs live v1).

Scoring

How a rule decides

Fire the rules, watch the score move

Demo rules

ApproveReviewBlockDecline

Score 45 lands inReview

Each rule adds points; the total lands in a band. The preset is the deck's structuring example: three transfers just under the reporting threshold.Source: Cortex Sentinel pitch deck, p.19 (rule weights and bands) and p.8 (structuring example).

The triage gate

What the machine may close

This is the team's own layer. A model recommends, but a fixed, auditable gate decides, and it fails closed: if anything is uncertain, a person gets the alert. Try to get the coin-flip alert auto-cleared.

Team build

Three-lane triage on the backtest corpus

Backtest, sample data

All alerts100%
Auto-closed by the triage gate6.5%
True positives among the auto-closed0%
No real case escaped through auto-close.

Only alerts the gate could clear safely were auto-closed; the rest went to an analyst or were escalated.Source: Cortex Sentinel project story (team backtest on sample/seed data). Not production results.

Manual review load, baseline rules vs the gate

Backtest, sample data

100Baseline rules

81.6With the triage gate

-18.4%

At a held 90% recall the gate cut false positives by 20.3% and manual review by 18.4%, against baseline rules at 100% recall and 18.1% precision.Source: Cortex Sentinel project story (team backtest on sample/seed data). Indexed: baseline = 100.

The decision

Who is allowed to close an alert?

Chosen: Transparent rules with a written reason

A model recommends, but a fixed, auditable gate decides, and it fails closed: if anything is uncertain, a person gets the alert.

Cost I accepted: Only a small share of alerts is closed automatically: 6.5% in the backtest, on sample data.

Source: Cortex Sentinel pitch deck, p.15 (fails safe, guard chain) and p.19; project story (backtest on sample data).

What shipped

The console

01
Detect
Scenarios and rules run on every transaction, with versions you can replay for a regulator on any date.
02
Write a rule in plain English
The rule studio where the agent's draft lands for a person to review.
03
Watch the decisions
Every decision is logged with its outcome and score.
04
Model your own data
Transactions, accounts and Travel Rule messages are mapped once, so every rule and review reads the same source.