Jev (TypeSafe AI)
OpenAI API
Vercel
GitHub
Claude Design
01
Tidey Sorts the Morning
Tidey investigates every alert overnight, closes the clear false alarms and splits the rest into three groups.
01The Problem
Every alert starts with groundwork
Before anyone can decide, someone has to piece the story together: who the customer is, what’s normal for them, what changed and when.
From interviews with fraud teams in North America and Europe (NICE Actimize, 2026). NICE Actimize sells fraud software, so these are vendor figures.
02The Idea
Split the decision
Code does the exact work, Jev makes the judgment calls with probabilities, an AI model writes for people, and a person has the final say.
-
≥ 0.85
Escalate as Fraud
Any fraud type at 0.85 or higher goes to the fraud team with a recommendation.
Fraud to Approve -
The rest
Send to a Person
Anything that misses a rule, with a summary and a suggested next step.
Judgment Calls -
≥ 0.85
Close It Alone
Legitimate at 0.85 or higher, nothing unexplained above 0.35, scam risk under 0.6 and under $5,000.
Closed by Tidey
03The Analyst’s Side
What the analyst sees
The queue says what’s waiting and where to start, and every case ends in one decision bar.
02
Approve a Block
An account takeover. Tidey recommends a freeze, and nothing happens until the analyst approves it.
03
Make a Judgment Call
Mixed signals on gift card purchases. Tidey suggests contacting the customer, and the analyst sends the text.
04
Spot-Check a Close
Every alert Tidey closes keeps its evidence. The analyst checks it and confirms the close.
05
Check the Test
Every decision from the held-out test, side by side with a single AI model.
04Design Decisions
How far Tidey is trusted
Six choices shaped how it decides and where a person steps in.
Questions, not one decision
Probabilities make every line tunable
Jev answers narrow questions. Code combines the answers, so each threshold can be read, tested and moved on its own.
Rules in code, not prompts
Policy stays policy
Moving where Tidey closes an alert is a one-line change a fraud lead can review, not a prompt rewrite.
Recommend, don’t act
A person approves every block
Freezing an account hits a real customer and is hard to undo. Tidey recommends and explains; an analyst decides.
A hard cap
Never closed alone at $5,000 or more
However confident Tidey is, large amounts always reach a person. The single model gets the same guardrail.
Checks on demand
Jev picks the optional checks
Required checks always run. Jev decides whether the rest are worth running, which keeps each investigation short.
Write for people only
A summary only when someone will read it
Closed false alarms keep their evidence but skip the AI-written summary, which is most of the cost.
05The Test
Tuned on one set, tested on another
Two sets of 50 synthetic alerts, each with an answer key neither version sees. Tidey was tuned on the first, then both versions ran once on the second.
Getting to 45Correct decisions out of 50
Development v1
31/50
My first question was the wrong one: ‘is everything explained?’ Almost nothing is fully explained, so 19 legitimate alerts went to a person.
Development v2
45/50
Turned around to “is anything out of character and unexplained?” One question, 14 more right.
One rule moved
48/50
Three legitimate alerts sat just over the 0.30 limit. Moving it to 0.35 replayed at 48 on the tuning set.
Held-out test
45/50
Run once on the second set. Nothing was re-tuned after it.
Held-out set
50fresh alerts
Generated after tuning, with a new seed, new places, merchants and amounts, and 8 new unclear cases.
24 legitimate18 fraud8 unclear
Scoring
One mistake costs more than the rest
- RightClose a legitimate alert, escalate fraud, or send an unclear alert to a person.
- CostlyA fraud or unclear alert closed with nobody looking.
- CheapA legitimate alert sent to a person for an extra review.
06Results
About as accurate, with different mistakes
Tidey trades 3 extra reviews for half the risky closes, at a third of the cost and 7× the speed.
| Measure | TideyJev decides, code acts | One AI modelGPT-5.4 mini, doing everything |
|---|---|---|
| Correct decisions | 45 / 50 | 46 / 50better |
| Fraud closed with nobody looking | 0 | 0 |
| Unclear alerts closed with nobody looking | 2better | 4 |
| Unclear alerts a person saw | 6 / 8better | 4 / 8 |
| Legitimate alerts sent to a person | 3 | 0better |
| Right fraud type when escalated | 18 / 18 | 18 / 18 |
| Cost per alert | $0.0012better | $0.0037 |
| Time to a decision | 0.30 sbetter | 2.2 s |
| Broken or invalid outputs | 0 | 0 |
Official run, October 14, 2026, with Jev (jev-1.13.0) and OpenAI GPT-5.4 mini. Tidey’s cost includes the summary it writes for alerts a person reads.
07Human Oversight
Built to hand off
The points where a person always decides, and the two alerts that got past them.
- Every block or freeze Analyst It hits a real customer and is hard to undo. Tidey only recommends.
- Judgment calls Analyst Mixed signals or scam risk. Anything that misses a rule goes to a person.
- $5,000 and over Analyst Never closed alone, however confident Tidey is.
- Closed alerts Analyst, spot-check Every close keeps its evidence. Reopen sends it for a second look.
- The lines themselves Fraud lead Thresholds are policy, set and reviewed in code.
The two misses
Both passed every rule. The single model closed them too, plus two more unclear alerts.
A-1036
$600 ATM withdrawal in Cancun
A booked trip, but a non-bank ATM at 2:40 AM is a common skimming pattern.
Legitimate1.00
Unexplained0.08
A-1049
$64 at a usual bookstore
From a new phone that took three tries at two-factor while the old one was still in use.
Legitimate0.99
Unexplained0.10
How the decision was madeA-1036
Who most likely did this?
Legitimate1.00
Stolen card used online0.00
Account takeover0.00
Card testing0.00
Anything out of character the evidence doesn’t explain?
Yes0.08
Could the customer be the victim of a scam?
Yes0.04
Code rule: Legitimate 1.00 ≥ 0.85 and nothing unexplained (0.08 ≤ 0.35), so Tidey closed it alone.
Answer key: unclear, a person should look. This decision is wrong.
Decided in 0.29 s for $0.0002.
The lesson
The question I didn’t ask.
Both misses passed every rule. Jev answered the question it was asked, and for both the customer was a fair answer. The fix is a new question, not a stricter number.
AskedWho most likely did this?For both misses, the customer was a fair answer.
NextIs this pattern risky even if the customer can explain it?A non-bank ATM at 2:40 AM, or a new phone with failed two-factor, still needs a person.
08Next Steps
What I’d do next
-
Pattern rules
A third question for Jev: “Is this risky even if the customer can explain it?”, for patterns like overnight cash at a non-bank ATM or a new device that failed two-factor. A high answer goes to a person.
-
Re-test
Run a fresh held-out set, so the new rules are checked on alerts they weren’t written for.
-
Shadow mode
Calibrate the lines on real labeled alerts, run next to analysts, and measure the minutes saved per alert.
Go Deeper
See it in action
A self-initiated project with synthetic data, built end to end: agent design, analyst app and a head-to-head test.