SafeSense

Pause. Check. Then act. A calm place to practice spotting scams, and to ask when something doesn't feel right.

Live: https://diejor.github.io/hackathon/ (works on a phone; no sign-up). HackUTD 2026.

What it does

SafeSense is for people who feel unsure about scams, often older adults. No timers, no pressure, a text-size control, and plain language throughout.

Part	What you do
Fraudle	Guess the scam word from a clue in six tries; every word comes with a one-line lesson.
Scam or Not?	Realistic texts from made-up companies: decide, then see the signs. Some are genuine, so it also teaches not to panic.
Role Reversal	You play the scammer. Pick tactics, watch a careful person take them apart, then a debrief turns each trick into a defense.
The Phone Call	Unlocks at 2 points. SafeSense rings your phone with a practice scam call (Retell voice agent). Private beta: invited numbers only.
Chat with us	The SafeSense assistant (Google Gemini) checks a message you type, say or screenshot: a verdict, the signs, what to do next. It remembers what you miss in the games. It's an AI and says so: guidance, not guarantees.
Text SafeSense	Forward a suspicious text over iMessage (Photon) to +1 (628) 316-6196 and get the same verdict back. Text PRACTICE for a labeled fake scam to practice on; reply REPORT if you spot it.

Points: every finished game is +1. A practice text or call is the real test: catch it for +1, get fooled and lose 1 (never below 0). Once a game unlocks it stays unlocked. No accounts: a recovery key on the progress page carries your progress to another device, and LINK <code> connects your phone's texts to the same memory.

SafeSense is meant as a companion to T-Mobile's Scam Shield: Scam Shield blocks at the network; SafeSense teaches people and covers what reaches them anyway. Not affiliated with or endorsed by T-Mobile.

Hackathon tracks

Track	How SafeSense addresses it
T-Mobile Safe Signal	Practice instead of fear, plus a second opinion on anything that slips past network filtering.
Best Use of Photon API	Signed Photon 2.0 webhooks handled inside the database, replies in ~3 s, LINK codes, opt-in practice texts graded by reply (REPORT = caught), STOP/START, caps and a kill switch.
MLH Best Use of Gemini API	Gemini reads text and screenshots, answers in a strict JSON schema, coaches from each person's game history, treats pasted content as data (prompt-injection rules), and falls back from flash to flash-lite.

Architecture

 Phone or laptop browser                               iPhone (Messages)            Phone (voice)
 safesense/ (React 19 + Vite, GitHub Pages)                  │ text SafeSense              ▲ practice call
   games ── log_activity ──┐                                 ▼                             │
   chat ──── bot_send ─────┤ WebSocket          Photon 2.0 ── signed webhook ──┐     Retell AI agent
   Call me ─ start_challenge_demo ┐                                           ▼          ▲      │ signed result
                           ▼      ▼                                                      │      ▼
        SpacetimeDB on Maincloud, db safesense-1aa18263 (TypeScript module, spacetimedb-ts/)
          private tables + per-user views · procedures with outbound HTTPS · scheduled jobs
          HTTP routes: /photon · /challenge-response · /voice/context · /voice/result
                           │
                           ▼
              Google Gemini API · Photon REST API · Retell API

There is no other server: the database is the backend. Identity is the SpacetimeDB identity whose token is kept in the browser (that token is the recovery key).

Folder	What's in it
safesense/	The app: React 19 + Vite, hash router. Emily's design; our wiring in src/lib, src/components, src/content
spacetimedb-ts/	The backend module: bot.ts (assistant, memory, Forget me), ai.ts (Gemini), activity.ts (game history), photon.ts + practice.ts (iMessage, practice texts), challenge.ts + voice.ts (practice calls), crypto.ts, schema.ts
web/src/bot/	Assistant data layer (connection, subscriptions, voice input) used through safesense/src/lib/stdb.ts
docs/	PLAN.md, PHOTON.md (iMessage), VOICE.md (calls), PITCH.md, DEVPOST.md, HANDOFF.md
tools/	phone.mjs (headless phone tester), photon-sim.mjs, voice-sim.mjs, call-sim.mjs (signed-webhook simulators for local testing)

Safety invariants

The texts and calls reach real phones, so these are enforced in the backend, not just promised:

#	Rule	Where it lives
1	Calls only go to invited numbers (CHALLENGE_ALLOWED_NUMBERS), with a 100-a-day safety cap	challenge.ts startChallengeDemo
2	Practice texts only go to a phone that texted the line first, max 1 a day (3 a day before linking)	practice.ts
3	Kill switch: nothing proactive goes out unless PHOTON_SENDING_ENABLED=true	photon-guards.ts
4	STOP / START honored with one confirmation, 30 replies per phone per day	photon-guards.ts
5	Fictional brands, .example domains, 555 numbers, no links; every practice text ends with a "SafeSense practice" label	practice.ts templates
6	No free AI text on the send path: practice texts come from fixed templates	practice.ts renderPractice
7	Every webhook is signed and checked (Photon HMAC, Retell signature), deduplicated, and rejected after 5 minutes	photon.ts, challenge.ts, voice.ts
8	No phone numbers stored at rest (only salted hashes); API keys only in the database environment, never in the website	voice.ts, challenge.ts, schema.ts
9	Private tables; each person sees only their own rows; screenshots are never stored; Forget me deletes everything	schema.ts, bot.ts

Run it locally

Prerequisites: Node 20+, the SpacetimeDB CLI 2.14.x, and a Gemini API key.

spacetime start --listen-addr 127.0.0.1:3001     # local server (CLI nickname `local` points here)
scripts/set-key.sh AI_API_KEY                    # hidden prompt; saved to the gitignored .env
scripts/publish-ts.sh                            # builds spacetimedb-ts/ and publishes db safesense-local
cd safesense && npm ci                           # also installs web/ (the site imports web/src/bot)
VITE_STDB_SERVER=ws://127.0.0.1:3001 VITE_STDB_DB=safesense-local npx vite

Just the website, no local server: point it at the live database instead (your browser gets its own identity there, so it's safe for clicking around; chat questions use the real Gemini key):

cd safesense && npm ci
VITE_STDB_SERVER=wss://maincloud.spacetimedb.com VITE_STDB_DB=safesense-1aa18263 npx vite

If the build fails with Property 'reducers' does not exist on type 'DbConnection', web/ is missing its packages: run npm ci in web/ (newer checkouts do this automatically).

Any build also accepts ?server=...&db=... in the URL. Other module settings (AI_MODEL, AI_FALLBACK_MODEL, PHOTON_WEBHOOK_SECRET, ...) are listed in docs/PLAN.md §5 and set the same way with scripts/set-key.sh NAME. Never commit or print secrets.

Deploy (project lead only):

• Module: STDB_TS_SERVER=maincloud STDB_TS_DB=safesense-1aa18263 scripts/publish-ts.sh (never -c on Maincloud).
• Site: scripts/deploy-safesense.sh builds safesense/ against the Maincloud db, keeps the old arcade under arcade/, checks the result and publishes it to GitHub Pages (--dry-run to only assemble). scripts/deploy-pages.sh now refuses to run without an explicit build dir, so it can't replace SafeSense with the arcade by accident.

Testing tools:

• node tools/phone.mjs <url> <outDir> <profileDir> <port> wait:4000 shot:home ... drives headless Chromium as a phone (touch, typing, file upload, screenshots). Use a fresh profile dir per simulated person.
• node tools/photon-sim.mjs safesense-local --checks sends signed, unsigned, tampered, stale and duplicate webhooks to a local db; with a message instead of --checks it runs one bot turn. Use a fake whsec_ secret and never point it at Maincloud. Details in docs/PHOTON.md.

Facts we cite

• In 2024, people reported $470 million in losses to scams that started with a text. The top five were fake package delivery, fake jobs, fake fraud alerts, unpaid tolls and wrong numbers. FTC Data Spotlight, April 2025
• Malwarebytes tracked "more than 1,000 closely related campaign templates" in a carrier rewards-points texting scam, monitored since early May 2026 (no victim or loss count given). Malwarebytes, 2026-09-17
• In a smishing campaign targeting US telecom customers, the scammers used genuine carrier links, with copies aimed at AT&T and Verizon customers. Whalebone

Roadmap

1. Verify-your-number for calls: text a code to the number before SafeSense may call it, so The Phone Call can open beyond invited numbers.
2. Personalized calls: the voice agent already can fetch the tricks a person misses (/voice/context); wire it into the agent's prompt.
3. Server-side points awarded in log_activity, so the web, iMessage and calls count the same, and history trimming never drops points.
4. Optional sign-in (SpacetimeAuth), with "bring my progress" from the anonymous key.
5. A dedicated Photon line, so anyone can text SafeSense, not one registered phone.
6. Partner insights: anonymous aggregates (which tricks still work), groups under 5 hidden.

Team and credits

Made by the SafeSense team at HackUTD 2026:

• Diego Rodrigues: project lead; assistant and backend.
• Krishanth Sentil: game design (BadScam, now Role Reversal), scam research, practice texts and the Retell practice call (The Phone Call), the mascot.
• Emily Yuan: product design and UI (the SafeSense design).

Built with Claude Code. Powered by SpacetimeDB, Google Gemini and Photon. Icons: Lucide.

Legacy

Security Arcade, Help Bot (C#) and Unity client notes

Before SafeSense, this repo held the Security Arcade (multiplayer phone games in Phaser on a C# SpacetimeDB module, db hackathon-1aa18263, still live at /arcade/) and a Unity 6 + Rish client. The assistant started there as the "Help Bot" and was ported to TypeScript with the same names.

Everything still useful is in docs/LEGACY.md: the arcade's lobby and security model, the log_activity contract and bot.ts API, Discord setup, C# build scripts, and the Unity/Rish gotchas. The most common trap: the C# scripts publish the unoptimized wasm (-b) because system wasm-opt (binaryen v133) emits wasm the spacetime CLI can't parse.