1. Avatar Creation
   Designing personalized avatars was a surprising challenge. We needed them to load fast, be customizable, and stay lightweight. Initially, complex image assets slowed the UI.
   Solution: We switched to a modular SVG-based avatar system with lazy loading, keeping the experience smooth without sacrificing personalization.

2. Integrating LLMs for Mock Interviews
   LLMs were critical for generating mock interview questions and feedback. However, early versions produced questions that were too random or generic.
   Solution: We applied prompt engineering and role-conditioning, framing the LLM as an interviewer for specific tech roles. We also built a validation pipeline to manually review and refine outputs before production.

3. Using Fluvio for Localhost Streaming
   Using Fluvio for real-time event streaming (e.g., Screenpipe logs) was powerful but difficult to configure locally, especially across different OS setups.
   Solution: We containerized Fluvio using Docker, allowing any developer to set up a working environment with one command, solving compatibility and setup issues.

4. Segregating Candidate and Recruiter Dashboards
   Initially, handling two distinct user dashboards caused navigation bugs and session leaks.
   Solution: We implemented strict role-based routing and built micro-frontend modules for each dashboard, ensuring clean separation and easier updates.

Other Key Learnings:

Strong architecture early on (especially user role management) prevents scale issues.

Always manually validate LLM outputs before trusting automation.

Containerizing services early avoids “works on my machine” problems later.