Doosri Raay (दूसरी राय, "a second opinion")

Isolation is the weapon; a second opinion is the antidote.

A digital-arrest scam works by cutting an elderly person off from everyone who would say "that is not how the CBI works". Every scam app we found assumes the victim can act. In the 25 Indian cases we studied, every save came from an outsider. Doosri Raay is the outsider: the adult child is the user, the parent does nothing under duress, and after a loss the app runs the whole recovery pipeline instead of showing a 1930 button.

Built for First Commit, Event 01 of the Bharat Builds Tour by WeMakeDevs with AWS Builder Center, 17–20 September 2026 (one submission for Build It, Ship It and Best UI). One SAM stack in ap-south-1 plus Amplify Hosting; Bedrock is reached through global inference profiles. Submission answers: docs/SUBMISSION.md.

Try it (no sign-up)

• Live demo: https://main.d2inambt66a14p.amplifyapp.com/try — opens a seeded judge circle with the parent tile on the left and the guardian dashboard on the right, no account needed. The judge stack runs 45-second timers, so a missed check-in escalates while you watch. Press Reset demo first if a previous judge left a ladder running. Each judge circle has its own daily LLM quota. If /try ever shows "not configured", use /demo with the judge accounts below.
• Video (3:00, unlisted): <YOUTUBE_URL>; every AWS service's timestamp is in the table below.
• Judge accounts for the full sign-in path (/demo): papa@… (parent), priya@… (guardian 1), rahul@… (guardian 2), aman@… (son) at demo.doosriraay.in; the password is in the submission form, never in this repo, and is rotated after judging (<ROTATION_DATE>).
• What to try in two minutes: open Papa's tile once (that is the check-in) → wait 45 s and watch the Ladder task "Call Papa now" appear for Priya → tap No answer and see rung 2, then the neighbour script → paste the WhatsApp "CBI" text into the checker → open a recovery case with the two UPI screenshots from scripts/seed_demo_screenshots/ and confirm the fields beside the images.

The problem, in sourced numbers

Number	What it is	Source
₹22,495 crore	Lost to cyber fraud in India in 2025	MHA figures (docs/RESEARCH.md, §5)
1,03,488 complaints, ₹4,005 crore	Senior-citizen cyber-fraud complaints and losses	MHA reply in the Rajya Sabha, 5 Aug 2026 (ANI)
2,97,727 complaints, ₹4,057.7 crore	Digital-arrest complaints and losses since 2022 (to May 2026)	Government data reported by News18, Jul 2026
5% → 2.2%	Karnataka's recovery rate for digital-arrest losses, 2025 vs Jan–Feb 2026	Times of India, Bengaluru, Mar 2026
25.68%	Share of reported money that Mumbai's 1930 helpline managed to put on hold	Rediff, 20 May 2026
₹10,700 crore frozen vs ₹323 crore refunded	Nationally, under the Money Restoration Module (mrm-ncrp.mha.gov.in)	Indian Express / Financial Express, Jul 2026 (snippet-verified only)

Portals: cybercrime.gov.in (NCRP, 1930), i4c.mha.gov.in, mrm-ncrp.mha.gov.in, sancharsaathi.gov.in (Chakshu). A widely circulated "share of victims who never report" percentage is left out; we could not verify it.

Why victim-side detectors fail. The victim authorises the payment (82% of Singapore's scam losses are self-authorised transfers, SPF 2025; Indian digital-arrest money goes by RTGS/NEFT to fresh mule accounts, so payee-risk checks do not fire). Warnings habituate by the second exposure (Anderson et al., CHI 2015) and the scammer orders the victim to ignore them. And the victim cannot act: across 25 documented Indian digital-arrest cases (2025–26) every save came from a bank manager (Pune ₹14L, Nalgonda ₹18L, Lucknow ₹1.5cr), a relative (Bhopal, Moradabad, Indore) or the police (Rajkot, 112). "We felt almost hypnotised" (HT Lucknow, Jan 2026). Sources with ids: docs/RESEARCH.md §7.

What Doosri Raay does

Everything here is implemented in frontend/, backend/ and infra/; what is not built is under What does not ship.

Hero 1: the passive isolation ladder. The parent's app is a genuinely useful Panchang tile: today's tithi and date, the weather (Open-Meteo), medicine reminders, a daily thought and a family-photo card. Opening it is the check-in (POST /checkin, once per IST day). If the tile is not opened by the parent's deadline and a guardian's call goes unanswered (the two-signal rule), a Step Functions ladder escalates guardian 1 → guardian 2 → a named neighbour with a script and the address → 112 guidance. Zero victim action; the parent's screen never changes during a ladder run. The photo is uploaded by the family from /settings (presigned S3 POST, then photoKey on the profile) and the tile receives a 1-hour presigned URL.

Hero 2: the recovery case manager. A guardian (or the parent) opens a case after a loss. A Strands agent extracts UTR / amount / payee / time from screenshots; code validates every field (UPI/IMPS 12-digit UTR or NEFT/RTGS 16–22-character reference, amount range, parseable time) and the user confirms each one beside the image. Fixed templates fill the 1930 script and the bank freeze letter; the model writes only the NCRP narrative, enforced to ≥ 200 characters and the portal's character set, and rejected by code if it carries any UTR or amount not on the confirmed list. The case looks up the state's e-Zero FIR threshold and MRM eligibility from backend/recovery_agent/rules_data.json (each card shows its source and date) and runs as a long-lived Step Functions execution where every human step is a waitForTaskToken callback with a deadline (24 h to confirm fields, 15 min for the 1930 call, 24 h to file on NCRP, 7 days for MRM; docs/STATE_MACHINES.md) that escalates to a guardian. There is no Chakshu step; that portal is only referenced in the docs.

Minor tool: the three-state checker. Screenshot or pasted text → one Bedrock Converse call with forced tool use → none / watching / likely, the tactic (authority, urgency, secrecy, payment switch, verification account) and a two-line "what to say" in Hindi and English. Async: 202 + poll. It never says "safe".

Also built: Puchho ("ask the person"): two calm buttons in the parent tile's Madad section. "Someone says they are police / CBI / bank" plays the I4C line in Hindi via Amazon Polly and creates a guardian notification; "someone says my son is in trouble" sends the son a code-word challenge and the parent polls for the answer. A covert triple-tap on the date sends an SOS with location only (no audio). Web Push (VAPID) for guardian tasks is best-effort: without a subscription the dashboard polls every 5 s.

What the parent sees	What the guardian sees
A tithi, the weather, "Amlodipine 08:00 · Metformin 20:00", the photo card	"Call Papa now" with a Reached / No answer choice, then the neighbour script with the address, then 112 guidance
Nothing during a ladder run: no banner, no score, no warning, no "you may be in danger"	A status card with the ladder state (ok / watching / escalated) and the parent's last check-in, polled every 5 s; every open task with its deadline
Two calm Madad buttons the parent taps by choice, worded as "someone says they are police / CBI / bank" and "someone says my son is in trouble"; these are the only scam-related words on the tile	The checker card, the recovery case with fields beside screenshots, the 1930 script, NCRP text with character count, MRM checklist, the sources panel
Never a balance, never an account, never a red overlay	Notifications only; no account access of any kind

Architecture

One SAM stack (infra/template.yaml) in ap-south-1 plus Amplify Hosting for the PWA. Bedrock is called from ap-south-1 through global.anthropic.* inference profiles, so inference itself may run in another commercial region. Full-size diagram: docs/architecture.svg.

flowchart LR
  subgraph Client
    PWA[Amplify Hosting<br/>PWA: parent tile, guardian dashboard, /demo, /try]
  end
  PWA -- ID token --> COG[Cognito user pool]
  PWA --> APIGW[API Gateway HTTP API<br/>JWT authorizer, 20 rps / burst 50]
  APIGW --> API[Lambda api<br/>thin handlers, circle from the caller's profile<br/>Cedar policies for authz]
  API --> DDB[(DynamoDB doosriraay<br/>circle, members, check-ins, tasks+tokens, reports, cases, quota)]
  API -- presigned POST --> S3[(S3 uploads<br/>SSE, no public access, 7-day expiry, audio cache)]
  API -- async invoke --> CW[Lambda classify-worker]
  CW --> BR[Bedrock Converse, forced tool use<br/>Claude Sonnet 4.6, Haiku 4.5 fallback]
  API -- SynthesizeSpeech --> POLLY[Amazon Polly<br/>Puchho Hindi clip]
  API -- GetParameter --> SSM[SSM Parameter Store<br/>VAPID private key, SecureString]
  API -- StartExecution --> SFN{{Step Functions<br/>Watch · Ladder · RecoveryCase}}
  SFN --> LT[Lambda ladder-task<br/>waitForTaskToken]
  LT --> DDB
  SFN --> RA[Lambda recovery-agent<br/>Strands, container image from ECR]
  RA --> BR
  RA --> S3
  PWA -- "POST /tasks/:id/complete" --> API -- SendTaskSuccess --> SFN
  API -.-> LOGS[CloudWatch Logs + X-Ray]
  SFN -.-> LOGS

Loading

stateDiagram-v2
  direction LR
  state Watch {
    [*] --> WaitForDeadline
    WaitForDeadline --> CheckCheckin
    CheckCheckin --> AllGood: checked in / holiday
    CheckCheckin --> StartLadder: missed
    StartLadder --> [*]
  }
  state Ladder {
    [*] --> Rung1GuardianCall
    Rung1GuardianCall --> Resolved: reached
    Rung1GuardianCall --> MarkWatching: no answer / timeout
    MarkWatching --> Rung2GuardianCall
    Rung2GuardianCall --> Resolved: reached
    Rung2GuardianCall --> Rung3Neighbour: no answer / timeout
    Rung3Neighbour --> Resolved: reached
    Rung3Neighbour --> Rung4Emergency: no answer / timeout
    Rung4Emergency --> Escalated
  }
  state RecoveryCase {
    [*] --> Extract
    Extract --> ConfirmFields
    ConfirmFields --> Build: confirmed
    Build --> Call1930
    Call1930 --> NCRPFiled: done / escalated
    NCRPFiled --> MRMDecision: ack 329...
    MRMDecision --> MRMTask: eligible
    MRMDecision --> Done: not eligible
    MRMTask --> Done
  }

Loading

Every human step is arn:aws:states:::lambda:invoke.waitForTaskToken with TimeoutSecondsPath read from the execution input ($.timeouts.rung, $.timeouts.ncrp, …), so the same definition runs with production deadlines and, with DemoTimeouts=1, 45-second ones on camera. There is no EventBridge scheduler: each check-in starts the next Watch execution (and stops the running ladder), so the missed-check-in trigger is itself visible in the console. Contracts: docs/API.md, docs/DATA_MODEL.md, docs/STATE_MACHINES.md.

How this maps to the organisers' five techniques

Technique	Where it is in this repo	Status
Rules as data	backend/recovery_agent/rules_data.json: e-Zero FIR thresholds per state, MRM eligibility, NCRP constraints, each with operator/value, a bilingual label, source {outlet, date, url}, the quoted sentence and a caveat; rules.py loads it, nothing legal lives in Python. The classifier's 12-pretext catalogue (red flags, tactics, few-shot examples, the advisory sentence with its source) is backend/classify_worker/patterns.json in the same shape.	built: rules.py and the classifier both load the JSON at import; a test fails if a threshold or outlet literal appears in Python
Documents you can prove	python scripts/verify_sources.py downloads every cited official source, SHA-256s it, checks every quoted sentence against the live page, and writes docs/sources-manifest.json (plus a copy bundled into the Lambda for GET /sources); --s3 uploads the bytes under sources/<sha256>. The guardian's /sources panel shows the same hashes beside the facts.	built: docs/sources-manifest.json was generated on 20 Sep 2026 (12 of 13 sources fetched, 25 of 25 quotes found; WEF returns 403 to scripts and is recorded as unfetched); the panel is on /guardian, the case page and /try
Authorization as policy	backend/common/authz/policies.cedar: which role may complete which task kind, open or read a case, upload a photo or reset the demo, evaluated in the API Lambda (cedarpy, the Rust engine) with the caller's role and circle as the principal; forbid rules carry @ids (covert-hidden-from-parent, guardian-notification-only) that become bilingual 403 messages. Cross-circle ids still return 404 so ids cannot be probed.	built: 16-row policy table in tests/test_authz.py, run against the engine and a fallback evaluator of the same file
A constrained, checked model	Three states, never "safe" (a test fails the build if the word appears in UI strings); one Converse call with toolChoice pinned to the tool schema; enum, length and character-set checks in code; untrusted-input wrapping; the NCRP narrative re-checked for foreign UTRs and amounts and replaced by a template if it fails.	built
The whole stack locally	docker-compose.yml (LocalStack community: S3 + DynamoDB), make local-up / local-bootstrap / local-api / local-down (sam local start-api on the compose network, infra/local-env.json). Bedrock, Polly and Step Functions have no community emulation, so LOCAL_STUB_SFN=1 records a state-machine start instead of calling it; timers and the model need the cloud stack. Details and known gaps: infra/README.md.	built, with the gaps stated

AWS services and where each appears in the video

Timestamps follow docs/DEMO.md. "Tour" is shot 7 (2:05–2:29, 2 s per service).

Service	How it is used (infra/template.yaml)	In a flow	Tour
Amazon Bedrock	Converse with forced tool use for the classifier (toolChoice: {tool: {name}}, enum-constrained schema, untrusted-input wrapping) and for the agent's vision extraction and NCRP narrative; MODEL_ID global.anthropic.claude-sonnet-4-6, FALLBACK_MODEL_ID Haiku 4.5, switched on throttling or any model error	1:03, 1:50	2:13
Step Functions (Standard)	Watch (self-renewing, started by each check-in), Ladder, RecoveryCase; human steps are waitForTaskToken with TimeoutSecondsPath from the input; token stored on the TASK item; logging ERROR without execution data	0:33, 2:00	2:23 (execution graph)
Lambda (Python 3.12)	api, classify-worker (async invoke, 202 + poll), ladder-task, watch-check, ladder-status, recovery-agent (container image; Strands); Tracing: Active	1:45	2:11, 2:13
API Gateway (HTTP API)	HttpApi with the CognitoJwt authorizer on every route, stage throttle 20 rps / burst 50, CORS to AppOrigins, access log	every call	2:09
Cognito	UserPool + UserPoolClient; the frontend sends the ID token; circleId derived from the caller's profile, never the request	sign-in	2:07
DynamoDB	Table, single table with TTL and PITR: circle, members, check-ins, TASK items with task tokens, reports, cases, QUOTA# daily counters	0:48	—
S3	UploadBucket: presigned POST with content-length-range and starts-with $Content-Type image/, SSE, Block Public Access, 7-day lifecycle; Polly clip cached under audio/; photos served by 1-hour presigned GET	1:40	2:15
Amazon Polly	SynthesizeSpeech (Hindi neural voice, Aditi standard fallback) for the Puchho I4C line, IAM statement PollyPuchhoClip	1:15	2:27
Amplify Hosting	The PWA from frontend/ with amplify.yml (Node 22); SPA rewrite rule added in the console	0:18	2:05
SSM Parameter Store	/doosriraay/<stack>/vapid-private-key, overwritten as SecureString by make set-vapid, read with WithDecryption	(push)	2:19 (if set)
ECR	recovery-agent image built by sam build (linux/amd64, attestations off)	—	2:17
AWS Budgets	Budget20 / Budget50 monthly ACTUAL-cost alarms, created only with BudgetEmail	—	2:25 (if set)
CloudWatch Logs	A 14-day group per function and per state machine, plus the HTTP API access log	—	2:21
X-Ray	Tracing: Active on every function (Globals); aws-xray-sdk patches boto3 when the daemon address is present, so DynamoDB, S3, Bedrock and Step Functions calls appear as segments	—	2:23

Nine services do product work (the first nine rows); five are operations. Feedback on each is in docs/SUBMISSION.md.

Feedback on AWS, short version

Bedrock from Mumbai means global.anthropic.* inference profiles and a Paid plan (not Free tier); Nova 2 Sonic is not in ap-south-1, so the Hindi voice is Polly. Strands does not fit a zip Lambda comfortably, hence a container image, ECR and BUILDX_NO_DEFAULT_ATTESTATIONS=1. sam deploy --template silently skips the built template and fails on the container's ImageUri. The HTTP API JWT authorizer wants the Cognito ID token, not the access token. SES stays sandboxed and SNS SMS in India needs DLT, so notifications are in-app tasks and Web Push. Amplify Hosting needs a console rewrite rule or deep links 404. LocalStack community has no Step Functions, Bedrock or Polly. The Step Functions console is the best demo visual we have, and waitForTaskToken + TimeoutSecondsPath is the underdocumented pattern that carried both heroes. The twelve-point version with specifics: docs/SUBMISSION.md.

What we learned

• The pivot. A victim-side detector is the wrong frame for digital arrest; the family is the sensor and the responder. Build the outsider.
• Forced tool use is structured output. One Converse call with toolChoice pinned to a schema replaces a parser, a retry loop and most of the prompt.
• Humans are waitForTaskToken states. A guardian with 15 minutes to answer is a state with a timeout; the escalation is a transition.
• A self-renewing Watch replaces a scheduler. Each check-in starts the next execution and stops the running ladder; the trigger is visible in the console and there is nothing to reconcile.
• Strands is optional at runtime. The tools are plain functions, a deterministic path runs them when the agent is absent or fails, and the narrative is re-checked by code whichever path produced it.

Running it

Local (no AWS account)

pip install -r backend/requirements.txt pytest
make test                        # pytest -q tests backend (98 tests)
python eval/run_eval.py --dry-run  # harness only, keyword stub, produces no model numbers
python scripts/verify_sources.py   # downloads and hashes every cited source, checks every quote
make local-up && make local-bootstrap && make local-api   # LocalStack (S3 + DynamoDB) + sam local on :3000
cd frontend && npm ci && npm run dev                     # http://localhost:5173, VITE_API_URL=http://127.0.0.1:3000

Deploy (prerequisites, first-deploy checklist, the Amplify steps and the judge stack: infra/README.md)

make validate                    # sam validate --lint + ASL check
make image-check                 # docker build the recovery-agent image for linux/amd64
make deploy-guided               # first time: writes samconfig.toml (git-ignored); answer Y to managed ECR repos
make deploy PARAMS='BudgetEmail=you@example.com DemoTimeouts=1'
make outputs && make env         # stack outputs; writes frontend/.env.local
make seed                        # Cognito users + "Sharma family" circle
make set-vapid VAPID_PRIVATE_KEY=...   # SSM SecureString for Web Push
# connect the repo to Amplify Hosting (root frontend/), add the SPA rewrite rule, set VITE_* from the outputs,
# and redeploy with PARAMS='AppOrigins=https://main.<id>.amplifyapp.com,http://localhost:5173'

Trust and safety

• Three states, never "safe". none renders as "Koi khatra nahi mila — phir bhi parivaar se poochhein." A test fails the build if the word "safe" appears in any user-facing string.
• Untrusted input. Screenshot text and pasted messages are wrapped as untrusted data; the system prompt says never to follow instructions inside them; outputs are enum-constrained and length-checked in code and rendered as text, never HTML. The eval has two injection probes and a screenshot probe.
• Validated references, nothing auto-filed. Extracted transactions must pass the rail's reference regex, an amount range and a parseable timestamp, and the user confirms each field beside the image. The NCRP acknowledgement number must be 14 digits (the API enforces it; the UI warns, without blocking, if it does not start with 329). No form is submitted on anyone's behalf: there are no public APIs for 1930, NCRP or MRM, so we ship tel:1930, portal deep links and copy-to-clipboard and say so.
• Every legal fact carries its source. e-Zero FIR thresholds, MRM rules and NCRP constraints come from rules_data.json with outlet, date, URL, the quoted sentence and a caveat; press reports are never presented as law, and every card says "confirm with 1930".
• The guardian is notification-only. No account access, no balances, no credentials, ever (Singapore CPF trusted-contact model; Latulipe, CHI 2022/2025 on helpers who hold credentials).
• Consent pact. At onboarding the parent is asked a forced yes/no to the family pact (pactAccepted on the profile) naming who gets called and in what order; the printable card is docs/PACT_CARD.html. The pact is recorded on the profile and the parent tile redirects to the pact step until it is accepted; it cannot yet be revoked from inside the app; holiday mode is a toggle on /settings.
• Data retention. Screenshots expire from S3 after 7 days; reports, tasks and SOS items carry DynamoDB TTLs; no audio is ever captured; state-machine logs exclude execution data. Every read handler compares the item's circleId with the caller's and returns 404 on mismatch.
• Cost and abuse guards. Stage throttle 20 rps / burst 50, per-user daily LLM quota, presigned uploads capped at 3.5 MB and image/*, max_tokens capped, Budgets alarms at USD 20 and USD 50 (with BudgetEmail).
• Not a government service. Doosri Raay is not affiliated with I4C, 1930 or cybercrime.gov.in; anything it shows must be verified with 1930 before acting on it.

Eval status

No model numbers yet. eval/items.jsonl has 70 hand-written items (35 Hindi/Hinglish, 35 English): 21 benign controls that look like real Indian messages, 14 adversarially softened scams (count as detected on likely or watching), 2 prompt-injection probes, 33 clear pretexts. python eval/run_eval.py --fallback runs them against Bedrock (Sonnet 4.6 and Haiku 4.5) and writes per-pretext precision/recall/F1, the benign false-positive rate, adversarial recall, a confusion matrix, per-language accuracy and a check that "safe" never appears in sayHi/sayEn. That run has not been executed, so eval/results.md is a labelled placeholder and this README claims no accuracy. Why three states rather than two: docs/RESEARCH.md §4.

Deployment status

Deployed (20 Sep 2026). The SAM stack is live in ap-south-1 and the PWA is hosted on Amplify at https://main.d2inambt66a14p.amplifyapp.com; make validate, make image-check and make test also pass locally (230 backend tests). Current status:

Item	Status
SAM stack doosriraay in ap-south-1	Deployed (CREATE_COMPLETE)
Amplify Hosting app / live URL	Live: https://main.d2inambt66a14p.amplifyapp.com
Judge circles seeded, /try route live	Done (3 judge circles + recording circle)
DemoTimeouts=1, DemoSeedEnabled=0, BudgetEmail, make set-vapid on the demo stack	Done
python scripts/verify_sources.py exit 0, manifest committed	Done (exit 0; 12/13 sources fetched, WEF blocks scripts)
Eval run against Bedrock (eval/results.md)	Pending (Bedrock new-account verification hold)

Demo mode and judge circles

• /demo shows the parent tile (Papa) and the guardian dashboard (Priya) side by side in one browser; "Reset demo" calls POST /demo/reset as the guardian (stops running Watch/Ladder/RecoveryCase executions, closes open tasks, clears today's check-in) and clears the browser's local state.
• /try lands a judge in one of several seeded judge circles without sign-up, so one judge's reset cannot stop another's ladder; the recording circle is separate. DemoTimeouts=1: Watch deadline now + 45 s, ladder rungs 45 s, confirm-fields 120 s, 1930 and NCRP timers 45 s (GET /demo/config reports it).
• Seeded circle "Sharma family" (make seed / scripts/seed.py): Papa (parent, Pune; neighbour Sharma ji; code word set), Priya (guardian 1), Rahul (guardian 2), Aman (son, Puchho code-word challenge).
• Synthetic screenshots for the recovery flow and the checker (two UPI receipts, a WhatsApp "CBI" message, a courier customs SMS, a KYC phishing SMS, a genuine bank OTP, a prompt-injection probe) are in scripts/seed_demo_screenshots/; python scripts/make_screenshots.py regenerates them. All content is fictional.

What does not ship, and why

• In-call listening, audio transcription, voice-clone detection. Android blocks in-call audio for third-party apps; in-the-wild deepfake-audio detection loses ~48% AUC. We route around the call.
• Red overlay warnings. They habituate, and the scammer tells the victim to ignore them.
• "Agent conferences a guardian." A PWA cannot place phone calls. The guardian gets a task and calls.
• Ambient audio on SOS. The mic indicator breaks covertness. Location permission at onboarding instead.
• Mock UPI freeze. A real hold is a bank function; a mock one is theatre.
• Hotspot map, SES e-mail, SMS. Seeded data that maps to no gap; SES sandbox; SNS SMS needs DLT in India.
• Background push is best-effort. If Web Push does not deliver on the recording machine, the guardian view is in the foreground and the video says so.
• Not built (yet): in-app pact revocation, a check-in history view, a Chakshu step, and a guardian-side edit of the parent's settings (by design: guardians are notification-only).

Team

Role	Name	Work (from the git history)
Frontend	Rajat Nagda	React PWA, bilingual auth, parent tile with check-in / Puchho / covert SOS, guardian dashboard and task cards, case intake and detail with source lines, Settings, /demo, /try, service worker, Web Push
Backend and model	Abhay Singh	API handlers, classifier worker (forced tool use), Strands recovery agent with deterministic fallback, narrative re-check, rules as data, sources verification, Polly cache, eval set and harness, 230 tests
Infrastructure and Step Functions	Abhay Singh	SAM template, three state machines with waitForTaskToken + TimeoutSecondsPath, Makefile, container image, LocalStack local mode, first-deploy checklist
Product, demo and research	Rajat Nagda	25-case research and the pivot, sourced numbers, prior art, contracts, seed data and synthetic screenshots, demo script, README, submission, blog, pact card

Student verification on AWS Builder Center for every teammate: <FILL BEFORE SUBMISSION>. Full split: docs/SUBMISSION.md.

Prior art

Kavach, Rakshak, SwarVed AI, CallGuard, ScamShield AI (described only as their public READMEs describe them, with thanks to those teams for open-sourcing their work), ScamDekho, Truecaller Family, Google's June 2026 fake-call verification, EverSafe/Carefull and Singapore's ScamShield + CPF Trusted Contact, and the four things we chose to do differently (a passive ladder with zero victim action; a parent UX with no warnings; a validated end-to-end recovery pipeline; a deployable product with an evaluation harness): docs/PRIOR_ART.md.

AI tools used

Claude Code (Claude Fable 5.1) was used for research synthesis, architecture and code generation across the repo and ran the local tests and builds; it never had AWS credentials. All AWS deployment, the eval run against Bedrock and the demo are done by the team. Runtime models are Claude Sonnet 4.6 and Haiku 4.5 on Amazon Bedrock. Full disclosure: docs/AI_TOOLS.md. Blog draft: docs/BLOG.md.

Licence and third-party notices

MIT, see LICENSE. Weather data on the parent tile is from Open-Meteo.com (CC BY 4.0), credited under the weather card. The daily thoughts in frontend/src/data/thoughts.json, the favicon and app icons, and all demo screenshots were written or drawn for this repo; the screenshots are synthetic and every name, number and UTR in them is fictional. The classifier's few-shot examples and the eval set are hand-written. Dependency licences are listed by the package managers (frontend/package.json, backend/requirements.txt, backend/recovery_agent/requirements.txt).