I got tired of reviewing 300+ CVs, so I built an AI Agent to do first-round interviews

Vadym Shramko — Wed, 03 Jun 2026 10:23:25 +0000

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!

How I stopped begging for structured bug reports (and automated it with AI)

Vadym Shramko — Wed, 03 Jun 2026 10:18:31 +0000

If you work in a software team, you know the pain. You ask Product Managers or Designers to fill out a detailed Jira template for bugs (Steps to reproduce, Expected, Actual). They do it for a week, and then go back to dropping a raw screenshot in the team chat with the caption: "checkout is broken".

As an engineer, you're left guessing the platform, the state, and the steps.

I got tired of fighting human nature, so I decided to build a bridge between the lazy chat screenshots and our strict Jira boards. Here is how I built a bot that does the translation automatically.

The Tech Stack
I wanted to build this fast, so I went with my go-to stack:

Next.js (App Router) for the API webhooks.

Supabase for keeping track of chat IDs and Jira connections.

Google Gemini Vision API for the heavy lifting (reading the screenshots).

Jira REST API for creating the actual tickets.

How the workflow actually looks now:

A PM finds a bug and forwards a screenshot to our Telegram bot with a minimal caption (e.g., "login button does nothing on iOS").

The Telegram webhook hits my Next.js endpoint.

I pass the image and the caption to Gemini Vision with a strict prompt: "Analyze this UI screenshot. Generate a bug report title, steps to reproduce, expected result, and actual result."

Gemini returns a perfectly formatted JSON.

The backend fires a request to the Jira API, creating a new Bug issue, attaching the original image, and applying the AI-generated text to the description.

The bot replies in the Telegram thread with the BUG-123 link.

Total time: ~15 seconds.

The Result
Devs get their perfectly structured tickets, and non-technical folks don't have to change their habits. The "garbage in, garbage out" problem is completely solved.

I realized this is a universal pain point, so I actually packaged this workflow into a small micro-SaaS called BotBridge (link in my profile).

Have you guys tried using Vision models for QA or triage workflows yet? Would love to hear how other teams are handling the "lazy bug report" problem!

DEV Community: Vadym Shramko

I got tired of reviewing 300+ CVs, so I built an AI Agent to do first-round interviews

How I stopped begging for structured bug reports (and automated it with AI)