Building a HIPAA-Compliant Chatbot with AWS Lambda & Bedrock

This content originally appeared on DEV Community and was authored by LAVANYA LAHARI NANDIPATI

As a software engineer working in healthcare, I’ve seen first-hand how patients want instant answers but compliance makes everything tricky. We can’t just throw PHI (Protected Health Information) into ChatGPT and call it a day.

So I decided to build a HIPAA-friendly chatbot using AWS Lambda (serverless backend) and Amazon Bedrock (LLM service) — with data masking to keep sensitive info safe.

This post walks you through:

• Why chatbots in healthcare are challenging
• How to design a compliant architecture
• Code snippets for AWS Lambda + Bedrock
• Tips for keeping PHI secure

The Challenge: AI + Healthcare = Risk

• Chatbots are great for FAQs, triage, and scheduling.
• But if you send raw PHI (like names, MRNs, diagnoses) to an LLM… that’s a compliance nightmare.
• HIPAA requires minimum necessary access and strict controls.

Architecture at a Glance

Flow:

1. Patient message → API Gateway
2. Lambda pre-processor → scrubs PHI (names, DOB, SSNs)
3. Bedrock LLM → processes the masked query
4. Lambda post-processor → reinserts placeholders if needed
5. Response → returned to patient securely

API Gateway → Lambda (mask PHI) → Bedrock → Lambda (restore placeholders) → Patient

Step 1: Setting up AWS Lambda

A basic Python Lambda handler:

import boto3
import re

bedrock = boto3.client('bedrock-runtime')

def mask_phi(text):
    # toy example: replace dates + names
    text = re.sub(r'\d{2}/\d{2}/\d{4}', '[DATE]', text)
    text = re.sub(r'\b(Alice|Bob|John)\b', '[NAME]', text)
    return text

def lambda_handler(event, context):
    user_input = event['queryStringParameters']['q']
    masked_input = mask_phi(user_input)

    response = bedrock.invoke_model(
        modelId="anthropic.claude-v2",
        contentType="application/json",
        accept="application/json",
        body=f'{{"prompt":"{masked_input}"}}'
    )

    return {
        "statusCode": 200,
        "body": response['body'].read().decode('utf-8')
    }

Step 2: Talking to Bedrock Safely

• Always send masked input only.
• Example:

Input:

John Smith has a fever since 09/21/2025. Should he see a doctor?

Masked:

[NAME] has a fever since [DATE]. Should they see a doctor?

Step 3: Post-Processing

If you need to restore placeholders (like Hello [NAME]), you can map them back safely from session state.

Why This Matters

• Patients get instant responses
• Engineers stay HIPAA-compliant
• Serverless (Lambda) keeps costs low
• Bedrock provides enterprise-grade LLMs without exposing PHI

Next Steps

• Add DynamoDB to store chat history (encrypted)
• Plug in Cognito for authentication
• Expand PHI scrubbing with Amazon Comprehend Medical

Closing Thoughts

As engineers, we often think “just ship the feature” — but in healthcare, privacy is the feature.

This project taught me that it’s possible to marry AI innovation with compliance if we design carefully.

Repo link: GitHub – hipaa-chatbot-bedrock

This content originally appeared on DEV Community and was authored by LAVANYA LAHARI NANDIPATI