If you’ve ever been involved in hiring developers, you know the drill. You open a position, and within 48 hours, you have 300+ PDFs sitting in your inbox.
Most of them are spam, heavily exaggerated, or completely irrelevant. But to find the hidden gems, your engineering team has to waste hours scanning PDFs and doing repetitive 15-minute screening calls just to check if the candidate actually knows what an API is.
It was draining our team's time. So, I decided to build a system that acts as an automated "first-line" technical recruiter.
Here is how I built an AI agent to parse CVs and conduct chat-based technical interviews before a human ever steps in.
The Architecture
I needed a stack that could handle file processing, real-time chat, and AI orchestration.
Next.js for the frontend and API routes.
PDF parsing library to extract raw text from uploaded CVs.
LLM (Gemini / GPT) acting as the core brain.
Supabase to store candidates, chat logs, and final scores.
How the AI Interviewer Works:
Context Loading: The candidate uploads their CV. The backend parses the PDF and feeds it into the LLM's context window.
System Prompting: The AI is given a strict persona: "You are a Senior Engineering Manager. Review this candidate's CV. Ask them 3-5 technical questions based specifically on the stack they claimed to know. Do not ask generic questions."
The Chat: The candidate enters a chat interface. The AI greets them and asks a specific question (e.g., "I see you used React and Redux at your last job. Can you explain how you handled state normalization?").
Dynamic Follow-ups: Unlike a static form, the AI analyzes the candidate's answer. If they give a vague ChatGPT-like response, it pushes deeper.
The Handoff: Once the chat is done, the AI generates a short summary and a "Match Score" for the human hiring manager.
The Result
Now, instead of jumping on 20 calls, our team only talks to the top 3 candidates who actually proved their knowledge in the AI chat.
It worked so well internally that I polished the UI and launched it as a standalone tool called Ghost-assistant (you can check it out at https://ghost-assistant-app.vercel.app/).
Building AI agents that actually interact dynamically instead of just summarizing text has been a super fun challenge. Has anyone else experimented with building AI interviewers or HR automation? Would love to hear how you handle prompt injection in these scenarios!
Top comments (0)