Plainly in 90 seconds
Six links, in order. The samples open instantly because their results were prepared in advance. Step 4 is a live check: your browser reads a photo and only the words are checked.
The tour
-
A scam that looks official 15 s
An electricity "final notice" demanding payment to a personal UPI ID tonight. Look for the red verdict, each warning sign quoted from the letter, and where to report it.
-
A notice in the genuine format 15 s
An IRS balance-due notice whose contacts match the official list. The verdict still sends you to the official number. The deadline is worked out in code from the letter's own date (on a live check, "Add to calendar" saves any date still ahead). Then switch the explanation to हिन्दी at the bottom.
-
A letter that talks to the AI 10 s
This letter hides a faint line meant to sway an AI checker. Plainly treats such text as a strong scam sign, and the result never repeats it. This sample's result was worked out from the letter's full text, faint line included. The reader on your device may not pick up faint print from a photo, so uploading this image in step 4 may not show the flag; the rule catches instructions that reach it as text, such as pasted messages, emails and PDFs with a text layer.
-
A photo, read on your device 20 s
Save the IRS sample image (or use any photo of a letter), then choose it on Check a letter. A progress bar shows Tesseract reading it inside the browser; the photo is never uploaded. The words appear in a box to correct, and only that text goes to the rules engine on AWS Lambda.
-
The receipts 10 s
On any result, open "Receipts: every check we ran" to see each step marked FLAG, PASS or UNKNOWN with its timing. On a live check the first line is the reading done on your device. The rules and the verdict logic are listed in full on How it works.
-
The build agent's receipts 10 s
The coding agent was connected to AWS through the AWS MCP Server under a least-privilege policy: no deletes through MCP, and no way to stop or reconfigure the audit trail once it was running. CloudTrail shows what it did.
Where each criterion shows up
| Criterion | What to look at | Where |
|---|---|---|
| Technical Innovation & Originality | The letter is read on the device, so the photo never leaves it. Plain code decides the verdict with 21 rules, each tied to a public source, and every check leaves a receipt. Hidden instructions aimed at AI tools are detected and never echoed. No "safe" verdict exists. | Verdict logic, receipts on any result |
| Implementation Quality | Serverless on AWS Free plan services: API Gateway, Lambda and DynamoDB, with one Lambda serving both the pages and the checks (CloudFront and S3 are ready in the template, waiting for AWS to verify the new account), and Tesseract.js self-hosted and loaded only when a photo is chosen. Amazon Textract and Amazon Nova are built and tested but switched off, because the Free plan doesn't include them; one setting turns Nova back on after an upgrade. Least-privilege access, per-address and daily limits, and Plainly stores nothing about a letter. | Evidence, Privacy, Textract and Nova |
| Community / Market Impact | Built for families handling official mail in a second language, in India and the United States. Pre-written explanations in English, हिन्दी and Español (nothing generated per letter), with a simple-words option. Hindi letters can be read on the device too. | Check a letter |
| Creativity & Storytelling | Two letters that look alike get opposite verdicts, and every claim carries its source. The agent that built the checker left receipts, just like the checker does. | The three examples, Evidence |