Scam Guardian

Pitch deck

PDF: ScamGuardianPitchDeck_ThreePeasInAPod_TaylorsUniversity

App screenshots

What Scam Guardian does

Scam Guardian demonstrates help for a person finding the transfer hard and a verification pause for a person who may be following someone else's instructions. Display categories come from demonstration policy, not a validated three-class detector.

This repository is a prototype for Dormathon 2026, Track 1, "Predict Early, Decide Better": an Expo application for Android and web with a product page and an interactive mock banking app. No real money moves.

Session	In the demo	Response
Normal	Saved recipient, usual amount	No interruption
Struggling	Three back steps or two help taps	Simple mode: one instruction at a time, with larger controls after repeated back steps
Coached	Call active and a new recipient	A 15-second pause that asks whether someone is directing the transfer
Coached, high	The above, plus a pasted account number or more than 3x the usual amount	The same pause, and a request for the trusted contact to call

After the pause the person can still continue. Guardian adds time and support; it never blocks someone's own money.

What is real and what is simulated

• Money, account and recipients: made up. No real money moves and no bank is connected.
• Call and app-switch signals: set by buttons in the demo. Nothing is read from the device's call state or from other apps.
• Account number: a real field taking 10 to 16 digits, typed or pasted. Guardian records how it was entered and the length of each change. The digits never leave the field.
• Classification: Normal, Struggling and Coached come from demonstration policy combining rules in src/domain/risk.ts and binary AI review.
• AI review: a frozen on-device model from the AI team scores the transfer every 5 seconds and can add a pause independently of the rules. It was trained on simulated sessions and its score is not a scam probability. See AI model integration.
• Family alert: a preview on a second phone drawn on the same screen. Nothing is sent to another device.
• Working as shown: the consent step, the separate trusted-contact consent, amount validation, Simple mode and the 15-second pause. The pause length is provisional, not a calibrated risk policy.

AI model integration

A frozen model bundle from the AI team runs inside the app, on the device or in the browser. Nothing is downloaded and no server is called. Its only output is binary: recommend a review, or do not.

The two models

Model	Used when	Inputs	Review threshold
xgb_behavior_binary (XGBoost)	No payment request is present	Behaviour only	0.724
lr_combined_binary (logistic regression)	An authored request has been delivered	Behaviour and request cues	0.952

Both separate Normal_or_Struggling from Coached. Struggling is never a reason for an AI review. The files, thresholds and routing come from src/domain/model/selection.json and are unchanged from the handoff.

How a transfer reaches the model

1. Collect. src/domain/collector.ts turns each action in an opted-in transfer into a content-free event: an edit with its length change and whether it was typed or pasted, focus and blur, navigation, help taps, and app or window focus. No digits, names or amounts are recorded.
2. Summarise. Every 5 seconds src/domain/features.ts reduces the events so far to twenty behaviour features, their change over the last 5 seconds, and the current step.
3. Encode the request. If the presenter delivers one of the three authored requests, src/domain/request.ts turns its text into eleven yes/no cues such as urgency, secrecy and "ignore the warning". This is an English pattern matcher on a fixed demo message, not language understanding, and never a real message from another app.
4. Score. src/domain/contextual_inference.mjs, the AI team's runtime kept byte for byte, picks the model and compares its score with that model's threshold.
5. Latch. The first score over the threshold latches a review for the rest of that transfer. Every prediction already due is awaited before Confirm is accepted.

What the app does with the result

• A latched review adds the 15-second pause, with the reason "Pause and verify this payment through a contact you already trust". The read-out shows whether the rules, the AI or both raised it. The AI never overrides a stronger rule response.
• If the bundle cannot be used, the read-out says the AI is unavailable. That is never treated as a safe result, and the rules still apply.
• When the logistic regression raised the review, the app shows the actual feature contributions behind that score, which rebuild it exactly. The XGBoost model has no such breakdown here.
• The policy category and priority in the read-out, and the suggested next step in the bank app's warning, come from a written display policy in src/domain/decision-support.ts. They are not learned outputs.

What the model does not see

The simulated call, the simulated app switch, motion, the amount and the recipient's history. Those can only drive the separate rules in src/domain/risk.ts.

See it in the demo

1. On /demo, pick Coached (the coached request with warning bypass) under Authored request in the controls card.
2. Press Transfer, choose the new recipient, then press Deliver.
3. Stay on the recipient step. The review appears at the next 5-second check, with Raised by AI beside the heading and the terms behind the score drawn as bars.

Without a delivered request, behaviour alone has not triggered a review on transfers made in this app: five app-shaped transfers scored on device stayed under 0.72.

How well it works

Evaluation	Coached sessions caught	Reviews on legitimate sessions
Original simulated test set	647 of 720 (89.9%)	2 of 720
New authored challenge, 32 sessions	6 of 16 (37.5%)	2 of 16

• In the original test, all 60 "quiet" scams, the ones with no telling wording, were missed.
• On the challenge, a hand-set baseline using only the request cues caught 8 of 16, so the AI is not shown to be better than simple cues.
• The behaviour-only model caught 0 of 16 on the challenge. Every label there reused the same behaviour, so that result says little about behaviour either way.
• Six of the ten misses have inputs identical to a legitimate session. No model could tell those apart from what it is given.

Full results are in the diagnosis. Reproduce them with node scripts/evaluate_pitch_evidence.mjs --verify.

Limits

• Trained and tested on simulated sessions only. Its score is not a scam probability.
• Training used 6 typed characters for the recipient, sometimes with 4 pasted. The app's account field takes 10 to 16 digits, which is outside that range. Retraining has been requested and has not happened.
• Request cues recognise a fixed list of English phrases. Reworded instructions are missed.
• Whether a warning changes what a person does has not been measured. An independent test set is prepared and not yet collected.

npm test checks that the bundled files match the handoff's hashes and that the app reproduces the AI team's worked examples. See reference/ai/README.md.

Run

Use Node.js 22.13+ and npm (validated here with Node 22.18.0):

npm ci
npm run web
npm run android

Android builds require Android Studio's SDK and Java 21, plus a running emulator or connected device. npm run android generates the ignored native project through Expo and compiles it.

Demo walkthrough

The product page at / only shows the bank app. Pressing the phone, or one of the four scenarios on the card beside it, opens /demo.

/demo shows three layouts: the Guardian read-out, Rohani's phone with the bank app, and Aina's phone as the trusted contact. On a wide screen they sit side by side and always fit the window. On a phone, a switch at the top moves between them.

Enable Guardian on the bank app's home screen. Family-alert consent is separate. Pick a preset under "Scenario preset", then start a transfer. The table at the top says what each one does.

A preset only sets the starting signals. Three of them can be changed by hand at any time:

• Start simulated call starts or ends the call.
• Simulate app switch shows a simulated other app and times how long it stays in front. During a monitored transfer the read-out shows this as "Simulated app switch". Neither the rules nor the model use it.
• A new recipient's account number is typed or pasted into the account field on the recipient step.

Verify and preview the application

npm run lint
npm run typecheck
npm test
npm run export:platforms
node scripts/preview.mjs

The application preview runs at http://127.0.0.1:4173, with the demo at /demo. Both Android and web bundles go into ignored dist/. These exports verify compilation, not APK assembly or device behavior.

Deploy

The web build deploys to Vercel as a static site with npm run build:web; vercel.json holds the settings. No secrets and no model server are needed. See deployment instructions.

More detail

• Model diagnosis: every missed case and the comparison with simpler baselines.
• Warning measurements: how lead time and false interventions are counted in the demo.
• Four-minute pitch and the /evidence page: the presentation script and its evidence.
• Manual tests: ten cases to run by hand.
• Validation: what was checked in each round, and what was not.
• Design notes: design decisions, sources and licences.

Project map

• src/app/: Expo Router landing and demo routes.
• src/screens/: responsive shared Android/web screens.
• src/components/bank/: the mock banking app, its transfer flow, the account-number field and the simulated other app.
• src/components/glass/: the frosted glass surface, as CSS backdrop blur on web and a translucent sheen on Android.
• src/components/guardian/: the read-out and presets, the product page's scenario card, the family phone, the phone frame, web beUI controls and platform art.
• src/domain/: the session reducer, demonstration risk rules, the signal collector, feature extraction, the on-device model bundle and tests.
• reference/ai/: the AI team's handoff files, kept byte for byte and used by the tests.
• tools/collectability-harness/: a separate Expo app that tests which behavioural fields can be collected. It is a test tool with its own package.json, not part of the product build.
• design/: the Hairline artifact and its source.
• docs/implementation-plan.md: research source, scope, remaining decisions and later integration work.
• AGENTS.md and CLAUDE.md: project instructions, including Conventional Commits.