Let me tell you about the three problems that made me build this.
First — every time a developer joins a new project, they spend days just trying to understand what they're looking at. No one has time to write good documentation, and no one has time to read bad documentation. Second — hiring managers are drowning in CVs where everyone claims to know React, Python, and "machine learning," but nobody can tell who actually writes real code versus who just copies tutorials. Third — AI coding assistants are powerful, but they live in a browser tab while your actual work lives in VS Code. You're constantly context-switching between the tool and the work.
I got tired of all three. So I built RepoLens.
What RepoLens Actually Is
RepoLens started as a GitHub repository scanner. It's now something significantly bigger — an AI system that onboards developers into codebases, screens technical candidates by reading their actual commit history, and operates as a VS Code agent that can open files, edit code, and run terminal commands on your behalf. Three distinct capabilities. One product. Built by one second-year CS student from the University of Debrecen, in his own time, with no team and no budget.
🔗 Live: repolens-mkn8.onrender.com
📦 GitHub: github.com/Inam-kakar
Feature One: AI Repo Onboarding
You paste any GitHub URL into RepoLens. It fetches the file tree, filters out the noise — node_modules, build artifacts, lock files — and sends the actual meaningful code to Groq's LLM for analysis. What comes back isn't a file list. It's a proper onboarding report: what the project does, how it's architectured, what the entry points are, what each major module is responsible for, and how everything connects.
Then — because reading is still work — RepoLens reads the report out loud to you through a voice assistant powered by ElevenLabs.
The reason I chose Groq over OpenAI for inference was purely speed. When you're scanning a repo with hundreds of files, every second of latency multiplies. Groq brought a 12-second wait down to under 2 seconds. That gap is the difference between a tool people actually use and one they open once and forget.
Feature Two: Candidate Intelligence — CV Screening That Actually Checks the Work
This is the part that changed everything about what RepoLens is.
Anyone can write "5 years of Python experience" on a CV. What they can't fake is their commit history.
RepoLens now takes a candidate's profile — their LinkedIn, their GitHub, their submitted CV — and does something no human recruiter has time to do manually. It visits their GitHub. It reads their repositories. It analyzes their commit frequency, the quality of their code, the complexity of what they've actually shipped, and how consistent their activity is over time. It then cross-references all of this against what they claimed in their CV and generates a complete candidate intelligence report.
That report scores them across four dimensions: credibility (does their experience match what they've built?), authenticity (is this their real work or copied?), visibility (how active and present are they in the developer community?), and technical depth (are they writing real logic or assembling boilerplate?).
For hiring managers, this means you stop interviewing people who look good on paper and start talking to people who demonstrably build things. For candidates who are genuine builders, it means their work finally speaks for them — not just their formatting skills.
Feature Three: The VS Code Extension — RepoLens as a Coding Agent
This is where RepoLens stopped being a tool and started becoming something closer to a collaborator.
The RepoLens VS Code Extension connects directly to your editor and gives RepoLens the ability to operate inside your development environment. It can open files. It can edit them. It can run commands in your VS Code terminal. You describe what you want — "refactor this function," "run the test suite," "find where this variable is defined and update it" — and RepoLens acts on it without you leaving your editor.
This is the Codex-style workflow that people have been asking for, but built natively into the RepoLens ecosystem so it understands the context of your project before it touches anything. It doesn't act blind. It already knows your codebase from the repo scan. That context is what makes the difference between an agent that does what you meant and one that does what you said.
The extension communicates with the backend via WebSockets for real-time bidirectional flow, and uses Groq's llama-3.3-70b-versatile for fast, context-aware responses that don't make you wait.
The Stack Behind It
The backend runs on Node.js and Express. The frontend is deliberately plain — vanilla HTML, CSS, and JavaScript — because I wanted zero abstraction between me and what the browser was doing. Groq handles all LLM inference. ElevenLabs handles voice. GitHub OAuth handles authentication for private repo access. Supabase and PostgreSQL handle users, reports, and storage. The VS Code extension communicates through WebSockets, keeping the connection live while you work.
Every choice in this stack was made for a reason. Speed, control, and the ability to debug deeply when things broke — and things always broke.
The Thing That Nearly Broke Me
Voice latency. Streaming audio from ElevenLabs in real time sounds straightforward until you realize that audio arrives in chunks, chunks don't align with sentence boundaries, and browsers have aggressive autoplay restrictions that kill your stream before the user even hears a word.
I spent over a week on this single problem. The solution was a combination of buffering at sentence boundaries so audio never plays mid-sentence, preloading the next segment while the current one plays so there's no gap, and using a user-gesture unlock on first interaction to get around the browser restriction entirely. It's not glamorous engineering. But it's the kind of problem that separates a product from a demo, and shipping through it taught me more than any course I've taken.
What I Actually Learned
I'm a 2nd year student. I have no senior engineer looking over my shoulder. Every architecture decision, every debugging session, every deployment failure — that's just me and the problem.
What RepoLens taught me isn't really about Node.js or LLMs or WebSockets. It's that the most important skill in building software isn't knowing the right answer. It's being willing to stay in the room with a problem long enough that the answer eventually shows itself. I stayed in the room a lot while building this.
The other thing it taught me: ship ugly. The version of RepoLens that people first saw was nowhere near what I imagined it would be. I held it back for weeks trying to make it "ready." The feedback I got in the first 48 hours after shipping taught me more than those weeks of polishing ever could have.
Try It. Connect With Me.
If you're a developer jumping into new codebases constantly, RepoLens will save you days. If you're a hiring manager tired of being misled by polished CVs, the candidate intelligence feature was built for exactly your problem. And if you're building in VS Code and want an agent that knows your project before it touches your files, the extension is where this is heading.
🔗 repolens-mkn8.onrender.com
I'm Inam Kakar — 2nd year CS student, 2 years of hands-on AI development, and genuinely obsessed with building things that solve real problems. If you work on developer tooling, AI infrastructure, or technical hiring, I'd love to talk.
LinkedIn: linkedin.com/in/inam-kakar-019931266
GitHub: github.com/Inam-kakar
If this resonated with you, leave a reaction — it genuinely helps other people find it. And if you've tackled similar problems or have thoughts on where agentic coding tools are heading, drop a comment. I read everything.
Top comments (0)