AI Career Coach Brings Personalized Guidance to Young People at Scale

For years, this nonprofit invested in young people through scholarships and programs — work that exposed a structural gap. Thousands of people ages 16 to 26 disengaged before reaching an application, and human-staffed support could only reach those who found their way to the door. The organization set out to extend individualized career guidance far beyond what a staffed model can deliver, placing an AI coach at the center of a purpose-built digital platform.

Protagona was engaged to launch a career development platform across web and mobile with an AI coach tuned for participants ages 16 to 26, integrated with education and mentorship partners, running on production-grade infrastructure with full observability, and equipped with analytics to support data-driven product decisions.

The coach had to be genuinely helpful on sensitive topics on career direction and economic anxiety, while remaining safe and on-brand for a youth audience. Guardrails required careful calibration: too restrictive, and the coach becomes useless; too permissive, and participant safety is at risk. The platform spans iOS, Android, and web, integrates with multiple third-party partners covering career Q&A, labor market data, continuing education, and mentorship matching, and must coordinate real-time AI responses with asynchronous background processing. Each partner introduces its own reliability and access dependencies. Foundation model selection added another layer of complexity: structured evaluation showed that cheaper, faster models could not meet the grounding accuracy required for live career coaching, making model choice a consequential product decision.

Protagona built an agentic AI coaching system using the Strands Agents SDK, hosted on Amazon Bedrock AgentCore and powered by Claude Sonnet 4.6. At runtime, the model decides which tool to invoke: recommending courses from a partner catalog, retrieving live labor market benchmarks, or handing the participant off to a structured career activity. A lightweight classifier routes conversations between the coach and structured activities at the top level. Session memory runs on Amazon Bedrock AgentCore Memory alongside conversation history in DynamoDB. The chat model is decoupled from the rest of the system, enabling configuration-based model swaps without redeployment, the same mechanism that allowed side-by-side evaluation before launch. A separate background agent runs nightly on its own AgentCore runtime to prepare recommendations and summaries.

Responsible AI controls were designed for the audience from the start. Bedrock Guardrails apply content filters, PII protection, and self-harm and crisis handling calibrated for ages 16 to 26. CloudWatch alarms track guardrail interventions, PII redaction volume, and chat latency. Human review gates partner education interest before any data export, and feature flags keep unfinished capabilities away from participants. Releases move through five environments with manual approval gates before staging and production.