
NeuroBridge: Rehearsing Hard Conversations with Amazon Bedrock, Serverless and Amplify
How a four-person student team built a low-pressure practice space for job interviews, salary talks and boundary-setting during the WeMakeDevs × AWS First Commit hackathon.
The Problem
Some conversations are hard before they start: asking for a raise, telling a coworker "no", phoning a clinic, walking into an interview. For people who find these moments stressful, including the anxious and neurodivergent users we designed for, "just practice" is good advice with nowhere to follow it. Friends aren't always available, and mistakes in front of a real person cost something. We wanted a place where the first attempt doesn't count.
What NeuroBridge does
A user signs up and picks one of eight scenarios: job interview, asking for a raise, talking to a professor, setting boundaries, asking for help, a clinic phone call, meeting someone new, or handling a product conflict. They then chat with an AI partner who plays a defined character, such as Taylor, an engineering hiring manager.
When the user finishes, they get written feedback on four dimensions (clarity, tone, responsiveness and composure), a short summary, what went well, and what to practice next. A progress page shows past sessions, a score out of 12 and a practice streak. Settings let people make the app quieter: text size, contrast, reduced motion, and how long the AI pauses before replying.

Architecture
Everything runs serverless in Asia Pacific (Mumbai) and is defined as code with AWS CDK.

- The React app is hosted on AWS Amplify Hosting.
- Users sign up and sign in with Amazon Cognito.
- The app calls Amazon API Gateway, which invokes AWS Lambda functions (Node.js 20, ARM64) for creating a session, fetching a session, sending a message and generating feedback.
- Creating a session loads the scenario's persona prompt and opening line from Amazon DynamoDB and stores the new session there.
- Each message goes to Amazon Bedrock through the Converse API, using Amazon Nova 2 Lite via a cross-Region inference profile. The Lambda sends the persona prompt plus the latest messages and stores the reply.
- Finishing sends the transcript to Bedrock again and asks for structured JSON feedback.
Service | What it does for us |
|---|---|
Amplify Hosting | Serves the React app over HTTPS |
Cognito | Sign-up, sign-in and email confirmation |
API Gateway + Lambda | The API, with one small function per action |
DynamoDB | Sessions and scenario prompts |
Bedrock (Nova 2 Lite) | The conversation partner and the feedback |
CDK / CloudFormation | The whole backend as versioned code |
Design decisions that mattered
- Scenarios are data. Each persona prompt lives in DynamoDB and is seeded by a script. Adding a scenario is a seed entry plus a card in the app, not a new Lambda.
- Predictable cost. Sessions are capped at 30 messages, replies at 300 tokens, and only the last 10 messages are sent to the model.
- The model gives ratings, not numbers. We ask for ratings and written notes, and the app derives the score from the ratings. The number always agrees with what the notes say.
- Comfort is a feature, not an afterthought. Motion, contrast, text size and pacing are user controls, and the interface respects the operating system's reduced-motion setting.
What went wrong, and what we learned
Our feedback praised nonsense. In testing we typed "NO" and "Bye", and got "You handled the conversation with composure and clarity" above four "Needs Practice" ratings. There were two causes. Our prompt didn't force the summary to agree with the ratings, and the app filled gaps with friendly default text. We fixed both:
- The prompt now requires the summary to agree with the ratings, never gives a "strong" rating unless the user showed the skill, and says honestly when there isn't enough input.
- If the AI call fails, the fallback makes no claims about the user.
- The interface only shows text the API actually returned.
The lesson: in a coaching product, a friendly lie is worse than an error message.
Deploying safely to a shared account.
- We ran
cdk diffbefore every deploy and read it ourselves, because by defaultcdk deploypauses only for security-related changes, not for replacing a table.
- We ran a CloudFormation drift check to confirm nobody had edited the stack by hand.
- We agreed that only one person deploys at a time, and we checked which AWS account our command line pointed at before every deploy.
Small gotchas that cost us time:
- API Gateway's "Missing Authentication Token" can mean a wrong HTTP method or path, not a missing login. Our
/pingroute only accepts POST.
- Amplify's single-page-app rewrite rule is JSON, and the regular expression's backslash has to be doubled.
- For a manual Amplify deploy, zip the contents of
dist/, not the folder.
Working with AI coding agents. We gave them a rulebook: plan first, feature branches, no secrets in the repo, AWS commands only with approval. Garbled characters like "â€" still showed up in the UI and docs. We fixed them at the source and added a UTF-8-only file rule.
Responsible AI
NeuroBridge shows a visible "practice tool, not therapy" note. The feedback never diagnoses or labels the user, and the wording avoids clinical or outcome claims. Conversations are sent to Amazon Bedrock and stored in DynamoDB [with an expiry: confirm the duration]. Progress (streak and XP) currently lives in the user's browser, so it doesn't follow them across devices.
What's next
Group-conversation scenarios, progress stored in DynamoDB (the table already exists), Google sign-in, sessions that survive a browser refresh, personalized scenario suggestions, voice with Amazon Transcribe and Polly, and production hardening such as stricter API access control and rate limiting.
Try it
Open the live demo , sign up with an email you can access, and enter the confirmation code from your inbox (check spam).
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article