Amigo

Amigo: The AI Interview Coach that listens. Built with Ragie & ElevenLabs, it turns static PDFs into a dynamic, real-time voice interview defense.

Comment

Here is the complete Hackathon Submission text, tailored to your story and the specific tech stack we built.

Inspiration

The idea for Amigo didn't come from a dataset; it came from a coffee shop conversation. I was sitting with a group of talented developer friends who were actively interviewing. Despite having strong GitHub profiles and solving hundreds of LeetCode problems, they were failing.

One friend confessed, "I freeze up when they ask about the 'Why'. I built a project using MongoDB, but when the recruiter drilled down into **why* I chose NoSQL over Postgres, or how I handled sharding, I panicked. I knew the answer, but I couldn't articulate it under pressure."*

We realized that current prep tools are broken. LeetCode tests your logic, and Resume Scanners test your keywords, but nothing tests your defense. We built Amigo to simulate that exact terrifying moment—when a recruiter looks at your resume and asks, "Tell me about this database choice"—so you can panic with an AI, not with your dream employer.

What it does

Amigo is an Audio-First AI Interview Coach that turns your static resume into a dynamic, voice-based interrogation.

  1. Contextual Ingestion: Users upload their PDF resume. Amigo reads it instantly using Ragie, indexing every skill, project, and bullet point.
  2. The "Phone" Interface: We built a custom "Phone UI" where the interview happens. It feels like a Facetime call, not a chatbot.
  3. Adaptive Questioning: Instead of asking "Tell me about yourself," Amigo finds specific details (e.g., "I see you used Supabase here. How did you handle Row Level Security?") and drills down.
  4. Real-time Voice Interaction: powered by ElevenLabs, Amigo listens, interrupts, and reacts with human-like latency.

How we built it

We focused on a "Composable AI" stack to keep the application lightweight and fast:

  • The Brain: We used Google Gemini 1.5 for the reasoning capabilities, allowing the persona to switch between "Friendly Coach" and "Skeptical Recruiter."
  • The Memory (RAG): We integrated Ragie for Retrieval Augmented Generation. Ragie handles the PDF chunking and vector retrieval, allowing the AI to cite specific parts of the user's resume without hallucinations.
  • The Voice: We utilized ElevenLabs Conversational AI via WebSockets. We connected ElevenLabs tools to our Ragie backend, so when the Agent asks a question, it is grounded in the retrieved documents.
  • The Backbone: Supabase handles authentication and maps the ephemeral voice sessions to the user's persistent document index.
  • The Vibe: The frontend is built with Next.js, styled with a "Vibe Coding" aesthetic (Glassmorphism, 3D assets) to make the experience feel premium.

Challenges we ran into

  • The "Phone" Illusion: Embedding the ElevenLabs widget inside a responsive CSS phone frame was trickier than expected. We had to use creative z-indexing and pointer-event manipulation to make the widget interactable while keeping the "iPhone bezel" overlay on top.
  • RAG Latency: Adding RAG usually slows down voice bots. We had to optimize our Ragie queries to ensure the AI didn't pause awkwardly for 5 seconds while "reading" the resume. We solved this by pre-fetching context at the start of the session.
  • Prompt Engineering the Persona: Initially, the AI was too nice. We had to aggressively prompt engineer the System Instructions to make it "interrupt" the user and be more skeptical, simulating a real high-stakes interview.

Accomplishments that we're proud of

  • Zero-Latency Feel: The conversation flows naturally. You can cut the AI off, and it reacts instantly.
  • The UI/UX: We are incredibly proud of the landing page. Taking the "Easy Trip" aesthetic and converting it into a functional AI tool makes the project feel like a real startup product, not just a hackathon script.
  • It Actually Listens: The coolest moment was when we tested it with a real resume, and Amigo asked a hyper-specific question about a niche library listed in the "Skills" section. It proved the RAG pipeline was working perfectly.

What we learned

  • Audio is the new UI: We learned that in voice apps, speed is a feature. Even a 1-second delay breaks the immersion.
  • Resume Data is Messy: parsing PDFs is hard. Ragie saved us hours of headache by handling the ingestion gracefully.
  • Panic is Universal: In testing, even we got nervous talking to Amigo, which proves that the "Simulation Theory" of interview prep works.

What's next for Amigo

  • Job Description Matching: Allow users to upload a specific JD (e.g., "Google Senior Engineer") so Amigo adopts that specific company's interviewing style.
  • Post-Interview Report Card: Generate a detailed breakdown of the user's "filler words," "confidence score," and "technical accuracy" after the call ends.
  • Multi-Modal Interviews: Bringing video avatars back into the mix for non-verbal communication analysis.

Built With

  • elevenlabs
  • knowledgebase
  • lovable
  • react
  • supabase
  • vercel
Share this project:

Updates