Powering Career Exploration With Multi-Agent AI
How we built a scalable conversational AI backend that guides high school students through team-based career simulations.
Educational platform
A US-based education platform helps high school students explore future careers through immersive, simulation-based learning. Students step into real professional roles, discover their interests, and build concrete action plans for their education.
eidos was brought in to lead the AI development: agent architecture, conversation orchestration, persistent context, evaluation, safety, and the infrastructure required to run the system reliably in production.
Challenge
What was needed
Students move through multi-step simulations, interact with different roles, make decisions, and reflect on the experience. The AI had to guide each interaction toward specific learning objectives while holding context and pedagogical consistency across the whole session.
At the same time, insights needed to evolve across multiple simulations, turning individual conversations into a broader picture of each student's interests and strengths.
One engine, three agent types
One shared, goal-driven engine — built with LangGraph and LangChain. 7 conversation configurations, 15 structured objectives, all coordinated in real time.
From prototype to production
Because the system serves students directly, safety was non-negotiable. Every message is sanitized before it reaches the model, PII redacted and prompt injection blocked, with a verification step confirming the filter actually worked.
Every conversation has been graded independently after the fact, on safety, instructional purpose, and topic drift, by a judge itself validated against human-labeled data.
The backend runs on FastAPI and AWS Bedrock, with conversation state persisted so students can pick up days later, and latency profiled stage by stage rather than guessed at.
Results
In February, an early pilot put students through complete, agent-guided simulations, from live decision-making to personalized reflection. The same architecture has run in production since then, handling hundreds of real student conversations.
Every conversation, then and since, was traced and independently reviewed for safety, feeding directly into each round of prompt and agent refinement.
Figures span the February pilot through September 2026. Eidos led development through May 2026; the architecture has run on its own since. Student counts are estimated, not a verified roster.
