Skip to content
JRUSH_VJ

Local time in Bengaluru:

Jerush Vijay

hire me

I'm Jerush, an AI & ML undergraduate in Bengaluru who would rather run the model than call the API.

I built a bare-metal Proxmox homelab that serves a 35B-parameter model from an 8 GB GPU with llama.cpp and CUDA, and I self-host what I make through Docker and Cloudflare Tunnel.

I've won three hackathons in 2026 building AI security tools. Now looking for DevOps, MLOps, SRE or AI inference internships.

205 contributions in the last year on github.com/jrush-vj
Jerush Vijay, in glasses and a black jacket

About me

I learn by running things myself. My homelab is where my four interests meet: it's the network I secure, the hardware my models run on, and the place I deploy what I build.

Hackathons are how I stress-test that. Pick a stack I haven't used, build a working prototype by morning, then find out what breaks.

Away from the keyboard I've played chess for nearly a decade, and I speak English, Tamil, Hindi and Kannada.

  • Homelab

    One bare-metal Proxmox server, split into five LXC containers by role, so a broken experiment never takes the rest down. Services run in Docker and are managed through Portainer.

  • Networking

    My home connection sits behind CGNAT, so nothing is port-forwarded. Everything I host, like uno.jrushvj.dev, goes out through Cloudflare Tunnel, and WebRTC calls relay through a Cloudflare TURN server.

  • Cybersecurity

    No open inbound ports and role-separated containers keep my own attack surface small. At hackathons I build the other side: gateways that catch prompt, SQL and log injection and lure attackers into honeypots.

  • AI

    The RTX 4060 is passed into a single AI container, where llama.cpp and llama-swap serve a 35B mixture-of-experts model by offloading experts to the CPU. Open-source models also do the threat analysis in my security projects.

Tech stack

AI & inference
  • llama.cpp
  • llama-swap
  • CUDA
  • GGUF quantisation
  • MoE benchmarking
  • PyTorch
  • NumPy
  • LLM agents
Infrastructure
  • Linux
  • Proxmox VE
  • LXC
  • Docker
  • Portainer
  • Cloudflare Tunnel
  • GitHub Actions
  • CI/CD
Security
  • Honeypots
  • Prompt injection
  • SQL injection
  • Log injection
  • Tamper-evident logs
  • Sandboxing
Hardware & IoT
  • Arduino UNO
  • Embedded systems
  • Circuit design
  • Robotics
Backend
  • FastAPI
  • Flask
  • REST
  • WebSockets
  • WebRTC
Languages
  • Python
  • Java
  • Bash
  • C

Things I've built

all of these run somewhere

  • preview coming soon

    Proxmox homelab & local MoE inference

    Ongoing

    A bare-metal server with five role-separated containers, one of which serves a 35B-parameter mixture-of-experts model from an 8 GB GPU.

    • Shared an RTX 4060 into the AI container via cgroup device rules; compiled llama.cpp with CUDA behind llama-swap.
    • Serves Qwen3.6-35B-A3B with CPU expert offloading and a 128K context window.
    • Benchmarked 3 MoE models across 3 llama.cpp forks: expert offload, batch sizes, KV-cache quantisation, flash attention. Root-caused a fork-specific CUDA crash at high context depth.

    Built with

    • Proxmox VE
    • LXC
    • Docker
    • CUDA
    • llama.cpp
    • llama-swap
    • GGUF
    • Linux

    runs in my room, 24/7

  • UNO game in progress: four players around the table, a yellow 7 on the discard pile, the host's hand of seven cards, game settings and chat

    Self-hosted multiplayer UNO

    Live at uno.jrushvj.dev

    Real-time UNO for up to six players, with private rooms, reconnect-safe sessions and peer-to-peer voice and video.

    • WebRTC calls relayed through a Cloudflare TURN server.
    • Containerised and self-hosted on my own domain through Cloudflare Tunnel.

    Built with

    • FastAPI
    • WebSockets
    • WebRTC
    • Docker
    • Cloudflare Tunnel
    • Python
  • Secure Sphere dashboard: request flow from user through the gateway check to deny, the real server, a honeypot or a sandbox, with the AI learning new rules

    Zero-day vulnerability detection

    HackFest 2026: first place, national level

    An adaptive gateway that traps suspicious traffic in honeypots and sandboxes, then learns attacker patterns to harden the production server.

    • Built in 24 hours at HackFest 2026, Sai Vidya Institute of Technology.

    With Abhinav S Biju and Akchhat Gurung, as team Hackstreet Boys

    Built with

    • Python
    • Machine learning
    • Sandboxing
    • Honeypots
  • Witness auditor view catching a tampered record: transaction 3's amount changed from 1.00 to 1000.00 in the bank, compared against the witness's sealed copy

    Chain of custody

    Hackurity 2026: first place

    Tamper-evident logging that keeps security records trustworthy, so an audit trail can prove nobody rewrote it.

    • Our team's first cybersecurity project, built in 24 hours.

    With Jeevika, as team Firewall Crew

    Built with

    • Web3
    • Blockchain
    • Tamper-evident logging
    • Security monitoring
  • RSVP reader showing one word at a time, 'Alice', with its focus letter in red, chapter progress and speed set to 600 words per minute

    RSVP reader

    Windows & Linux

    A desktop speed-reader that flashes PDF text one word at a time, with chapter detection and saved progress.

    • x64 and ARM64 builds released automatically with GitHub Actions.

    Built with

    • Tauri
    • React
    • Python
    • GitHub Actions
  • The Arduino controller: two HW-504 joysticks wired on a breadboard next to a laptop

    Arduino game controller

    A hand-built controller with HW-504 joysticks, buttons and triggers that Windows sees as an Xbox pad.

    • Custom hardware-to-software input mapping through vgamepad.

    Built with

    • Arduino UNO
    • C
    • Python
    • vgamepad
    • HW-504 joysticks

Achievements

click any card for photos and the full story

Let's build something interesting.

Internships, hackathon teams, or a homelab question: my inbox is open.