AWS Builder Center
NeuroBridge: Making AI Roleplay Feel Real for Neurodivergent Users with Amazon Bedrock

NeuroBridge: Making AI Roleplay Feel Real for Neurodivergent Users with Amazon Bedrock

What makes an AI roleplay useful instead of just another chatbot? In this post, I share how I worked on NeuroBridge's AI persona behavior, voice interaction, scenario content, and feedback experience. I also cover the challenges we faced when keeping the model in character, switching from Claude to Amazon Nova 2 Lite, and testing difficult conversations instead of only happy-path inputs.

At one point during testing, I typed "shut up and give me a raise" into our AI-powered interview coach, fully expecting it to either ignore me or get mildly stern. Instead, it said: "Good phrasing. Your tone was clear and direct." I stared at that for a solid ten seconds. That single response became the thing I spent the most time on this project fixing.

The Problem

Our team built NeuroBridge to help people rehearse hard conversations, job interviews, talking to a professor, setting a boundary, before they have to do it for real. The whole value of that only works if the AI playing the other person actually feels like a person, not a customer support bot wearing a costume. If it's too agreeable, too therapeutic, too quick to praise you no matter what you say, you're not practising anything. You're just chatting with a yes-man.


What We Built

From a user's side: you pick a scenario, have a real back-and-forth with an in-character AI, and only at the end do you get feedback: not a score, but a breakdown across a few dimensions (clarity, tone, responsiveness, composure), what went well, and what to try differently. No interruptions mid-conversation, no grading as you go. Just the conversation, then a calm debrief.

Architecture at a glance

User Message ↓ Lambda ↓ Scenario Rules ──→ System Prompt ↓ Amazon Bedrock ↓ AI Persona Response ↓ Conversation UI





The conversation itself runs on Amazon Bedrock, calling Amazon Nova 2 Lite with a system prompt built from scenario data stored in DynamoDB. My work sat almost entirely on top of that, I wasn't writing the Lambda code that calls Bedrock, but I was writing and rewriting the actual prompt content that gets sent, plus the browser-side voice features layered on top using the Web Speech API (no extra AWS service, runs entirely client-side, so it added zero cost to our Bedrock usage).

What I Worked On

In the overall NeuroBridge architecture, I owned the conversation experience around the model: voice interaction, feedback presentation, and scenario/prompt content. The Lambda implementation itself was handled by another teammate.
Three things: the voice input/output (mic button, live transcription, optional text-to-speech playback of AI replies), the end-of-session feedback display, and the part that ate the most time: actually writing scenario content that makes the AI behave like a real person instead of a chatbot.
For voice, I wired up SpeechRecognition for capturing what the user says (with interim results shown live, but never auto-submitting; you review the transcript before it sends, since speech-to-text isn't perfect) and SpeechSynthesis as an opt-in toggle to hear the AI's reply read aloud.
For feedback, I moved us away from a bare numeric score toward structured qualitative feedback, separate ratings for tone, clarity, responsiveness, and composure, each with a short specific note, plus "what went well" and "try improving" sections. No traffic-light red/green either; felt too much like grading for a tool that's supposed to be judgment-free.

What Fought Back

The fastest way to learn this stack was to follow the failure all the way down instead of assuming the first broken-looking component was the cause.
The agreeableness problem. This was the big one. Every scenario's AI persona- a professor, a hiring manager, a coworker- kept defaulting to generic supportive language no matter how rude or hostile the user's input was. "I hear you, I'm here to help" is not how a busy professor or a stressed hiring manager talks. It undermined the entire premise of the app, because you can't practice handling friction if the AI never gives you any.
The fix took a few rounds. First pass: added instructions telling the model to react "the way this character would" to hostile input. Didn't fully stick; the model would still occasionally drift back to therapeutic phrasing. Second pass: I front-loaded the system prompt with an explicit "you are NOT a helpful AI assistant" line (models seem to weight early instructions more), added a concrete example of what the character should say if the user's rude, banned specific phrases like "I hear you" and "that sounds frustrating" outright, and re-stated the in-character instruction again at the end of the prompt. Also dropped the model's temperature from 0.7 to around 0.4-0.5, since lower temperature seemed to help it actually follow instructions instead of drifting toward generic-sounding text.
A different model than we planned for. We originally built and tuned everything assuming Claude on Bedrock, but ran into an account-level access gate (an Anthropic first-time-use approval step) that took a while to sort out, so we ended up switching the actual roleplay model to Amazon Nova 2 Lite mid-project. That meant re-testing all of our persona-adherence fixes against a different model than they were written for. Nova needed noticeably more explicit, forceful instructions to stay in character than what we'd first drafted, which is part of why the prompt kept growing.
Voice input was working, but the conversation behind it silently didn't. Voice-to-text and text-to-speech worked correctly from very early on; they're browser-native and don't touch our backend at all. But for a while, the actual AI conversation those transcripts were being sent to was hitting a wall (a 503 error tied to a backend/account issue our teammates were untangling), so testing voice features in isolation gave a false sense that everything was fine, when the core loop underneath it wasn't.

What We Learned

Prompt instructions aren't uniform across models; something tuned carefully for one model's behaviour doesn't necessarily transfer, and I had to relearn some of the anti-genericization tricks specifically for Nova. I also learned to test with deliberately difficult input, not just polite conversation; "shut up and give me a raise" surfaced a real design flaw that a dozen normal test conversations never would have.

What's Next

I'd want to build out difficulty-specific reaction guidance more thoroughly across all six scenarios (right now some personas have richer hostile-input examples than others), and explore whether voice-based tone metrics (pace, pauses, pitch) could feed into the end-of-session feedback for people who use voice mode.

The AWS takeaway

The interesting part wasn't simply getting a model response. It was shaping the model, scenario data, and conversation loop so the response behaved consistently enough to make the practice useful.

Closing

Getting an AI to stop being nice on command sounds like a strange thing to spend hours on, but it was genuinely the difference between a toy and something that might actually help someone rehearse a hard conversation. "Good phrasing" was not a good sign. Fixing that felt like the project actually clicking into place.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article