Parlina Conversational Agent

Parlina Conversational Agent is a voice-first personal tutor embedded directly inside the Parlina learning platform. Instead of being a generic chatbot, it is grounded in a student’s actual course lessons, so it can answer questions, clarify concepts, and provide step-by-step guidance using the same materials the student is studying.

Technical Complexity

This project turns a learning app into a cohesive, multimodal agent system:

• Voice + real-time conversation via the ElevenLabs Agents Platform (low-latency interactive tutoring).
• RAG-based grounding using an automatically generated knowledge base built from each course’s lesson content, enabling accurate, course-specific explanations rather than hallucinated answers.
• UI action orchestration: when the agent references a lesson that contains deeper context, it can ask the student if they want to jump there and then trigger the app’s navigation flow—bridging natural language tutoring with concrete “take me to the right place” product actions.
• Lifecycle automation: course updates can sync the agent’s knowledge base, and course deletion can clean up associated agent resources, ensuring the system stays consistent over time.

Innovation & Creativity / Real-World Impact

Beyond “voice chat” alone, the innovation is voice tutoring that’s tightly integrated with learning workflow. Students can stay in flow: ask questions out loud, get explanations anchored in their curriculum, and immediately jump to the exact lesson that fills the gap. This reduces friction for learners who struggle with typing, attention, or confidence, and makes high-quality tutoring more accessible at scale.

Theme Alignment: browsers, voices, clouds, and tools → one agent

Parlina Conversation Agent achieves theme alignment by combining:

• Browser UI (Parlina’s course/lesson interface),
• Voice (ElevenLabs conversational experience),
• Cloud AI (agent reasoning + RAG grounding),
• Tools/actions (knowledge base sync, lesson references, and in-app redirection) into a single, coherent tutoring agent that can both converse naturally and drive meaningful in-product outcomes.

Spec-driven development process

Built with a deliberate, spec-driven workflow:

• Windsurf + GPT-5.2 Medium Reasoning Fast to analyze the repo, update developer documentation, and create AI steering guides for consistent implementation.
• OpenHands + Zed + GPT-5.2 Pro for deeper codebase analysis and an ~800-line implementation spec.
• OpenHands + Zed + GPT-5.2 to implement the spec end-to-end.

Prior work includes the working AI education platform with multimodal lesson content generation (rendered Markdown for “Reading”, infographic slides and audio voice-overs for “Lesson Player”), and adaptive evaluation.