Interview Arena

Inspiration

Preparing for an interview is notoriously difficult because feedback is usually vague. You might get a "no," but you rarely learn what you should have said. We realized that standard mock interview apps are one-dimensional. They score your answer, but do not show you a better one.

We asked ourselves: "What if you could watch an expert take the interview for you, using your own resume?"

This inspired Interview Arena, a platform that flips the script. Instead of putting the user on the spot, we created a "Spectator Mode" where AI agents compete to answer questions based on the user's actual work history. It turns interview prep from a stressful test into a strategic masterclass.

What it does

Interview Arena is an Agentic Training Platform that simulates a high-stakes interview environment.

The Judge: An AI Interviewer analyzes your uploaded resume and asks a challenging, context-aware question.
The Battle: Two distinct "Challenger Agents" (powered by LLMs) instantly generate optimized answers. They do not make things up. They pull real facts and metrics from your resume to construct the response.
The Verdict: The Judge blindly evaluates both answers, picks a winner, and explains the specific rhetorical strategies that made the winning answer superior.
The Insight: The user watches this battle unfold, effectively learning how to frame their own experience like a top-tier candidate.

How we built it

We built Interview Arena as a modern full-stack application, leveraging the latest web standards:

Frontend: We used Next.js 16 (App Router) and TypeScript for a responsive, server-rendered interface. The UI was styled with the brand-new Tailwind CSS v4, utilizing its new configuration-free engine to create a sleek "Glassmorphism" aesthetic.

Backend: The heavy lifting is done by a Python FastAPI server.

AI Architecture: We used the OpenAI Python SDK to orchestrate our multi-agent system.

Resume Parsing: We utilized PyPDF2 to extract raw text from user resumes, which is then fed into the context window for our agents.

Agentic Compare: We implemented a ComparatorService that runs parallel inference calls to generate conflicting viewpoints and then synthesizes a final verdict.

Challenges we ran into

Precision System Prompting: It was incredibly difficult to engineer the system prompts to make the AIs behave exactly as intended. We had to iterate constantly to ensure the "Challenger Agents" strictly adhered to the user's resume facts without hallucinating in the wrong direction, while simultaneously ensuring the "Judge Agent" remained impartial and critical. Getting them to stay in character required complex, multi-step instructions.

Finding a Purposeful Idea: Our biggest hurdle was not the code, but the concept. We brainstormed dozens of ideas but struggled to find one that felt like it solved a real human problem rather than just being a "tech demo." Landing on the "Spectator Mode" angle took time because we wanted to build a tool that actually lowered interview anxiety, rather than just another generic chatbot.

Accomplishments that we're proud of

True Agentic Comparison: We did not just build a chatbot. We built a system where models evaluate models. Achieving this "Agentic Compare" workflow was a huge technical win.

Personalized "Hallucinations": We successfully tuned the agents to "hallucinate" answers that are actually true to the user's life. Seeing the AI argue about my specific project experience felt incredibly magical.

A Modern, Fast Stack: Integrating FastAPI with Next.js 16 gives us the best of both worlds. We get Python's AI ecosystem and React's interactive UI.

What we learned

The Power of "Spectator Mode": We learned that sometimes the best way to teach is not to quiz the user, but to let them watch experts.

Multi-Agent Orchestration: We gained a deep understanding of how to manage state when multiple asynchronous AI agents are generating content simultaneously.

What's next for Interview Arena

Voice Integration: We plan to integrate ElevenLabs to give each AI agent a distinct voice (e.g., a stern Judge vs. an energetic Candidate). This will make the "Arena" feel like a real radio drama.

User "Tag-In": A feature where the user can pause the battle and "tag in" to answer the question themselves, challenging the AI's high score.

Industry-Specific Arenas: Creating specialized agent pools for Coding, System Design, or Behavioral interviews.