Inspiration
In small towns and villages around Durg, many people, especially elders, get electricity notices, bank letters and government forms they can't fully understand. Scammers take advantage of this with fake "your connection will be cut tonight, share your OTP" messages. A single misunderstood letter can cost someone their savings. We asked: what if anyone could just take a photo of a letter and have it explained out loud in their own language?
What it does
Saral Patra (सरल पत्र) is a voice-first AI helper:
- The user uploads a photo of any official letter or bill.
- It explains the letter in simple Hindi or English.
- It lists clear action steps, the deadline and the amount.
- It runs a scam check (LOW / MEDIUM / HIGH risk) with red flags.
- It reads everything aloud, so even people who can't read well can understand.
How we built it
- Amazon Textract extracts text from the uploaded photo.
- Amazon Bedrock (Amazon Nova) turns the text into structured JSON: document type, summary, steps, deadline, amount, risk level and red flags.
- Amazon Polly (Kajal, neural Hindi voice) converts the explanation to natural speech.
- Python + Streamlit + boto3 form the frontend and the pipeline that chains the three services.
$$\text{Photo} \rightarrow \text{Textract} \rightarrow \text{Bedrock} \rightarrow \text{Polly} \rightarrow \text{Simple voice explanation}$$
What we learned
- How to chain multiple AWS AI services into one working pipeline.
- Prompting an LLM to return strict, structured JSON that a UI can reliably display.
- Designing for non-technical users: big text, minimal steps and audio first.
- Handling IAM permissions and Bedrock model access.
Challenges we faced
- Getting the LLM to return clean JSON every time, so we parse the output defensively.
- Blurry or tilted phone photos reduce OCR quality, so we added a paste-text fallback.
- Keeping explanations in very simple Hindi instead of formal, difficult words.
- Building and testing everything within a tight hackathon deadline.
What's next
- WhatsApp bot, so users don't need to install anything.
- Chhattisgarhi, Marathi and more regional languages.
- PDF support and one-tap "report scam" to the cybercrime portal.
Built With
- amazon-bedrock
- amazon-nova
- amazon-polly
- amazon-textract
- amazon-web-services
- aws-iam
- boto3
- generative-ai
- ocr
- python
- streamlit
- text-to-speech
