The challenge
Active listening is one of the most important skills in therapeutic and coaching practice, but it's remarkably hard to learn well. The traditional approach requires supervised role-play sessions with experienced evaluators who can assess whether a response truly reflects, acknowledges emotion, avoids giving advice, and uses open questions. These sessions are expensive, hard to schedule, and impossible to scale.
Our client wanted to democratise access to active listening training — making it possible for anyone to practice and receive professional-quality feedback, any time, without needing a human evaluator present.
What we built
The platform operates in two distinct modes:
Listener mode. Users can have a conversation with an AI trained specifically in active listening techniques. The AI reflects back what the user says, acknowledges emotions, asks open questions, and never offers advice or solutions. It's designed to demonstrate what good active listening feels like from the receiving end — giving practitioners a lived experience of the skill they're learning.
Trainer mode. Practitioners select from five realistic client scenarios — work stress, relationship conflict, life direction uncertainty, caregiver exhaustion, workplace disappointment. They read the client's situation and write their response as if they were the listener. The AI then evaluates their response against six core criteria: reflection quality, emotional acknowledgement, non-directive approach, use of open questions, absence of minimising or premature reassurance, and appropriate pacing.
Each evaluation returns a score out of 10, specific strengths, areas for improvement, and a model answer showing what an expert response might look like. Practitioners can iterate on their responses and track their progress across scenarios.
Safety first. Because the platform deals with emotional content, crisis detection is built into every interaction. The system monitors for keywords related to suicide, self-harm, abuse, and psychosis indicators, and surfaces crisis resources (Samaritans, SHOUT) immediately when safety concerns are detected. The platform is explicitly positioned as a training tool, not a substitute for professional mental health support.
Architecture
The frontend is built with Astro and React using an island architecture — static pages with interactive React components hydrated only where needed, giving excellent performance. The backend runs two Lambda functions (listener and trainer) that communicate with Amazon Bedrock's Claude models. Authentication uses Cognito, and the entire site is served through CloudFront for global edge delivery.
The result
What previously required booking a session with a human evaluator is now available on demand. Practitioners can practice at midnight, get instant feedback, try again, and see their scores improve — all without scheduling constraints or per-session costs. The AI evaluation is consistent and calibrated against professional criteria, giving trainees reliable feedback they can trust.