system time
00 : 00
wed · 25 mar 2026
open to work AI · ML · SWE · DevOps
identity
Vishal
Katariya
Darmstadt · DE h_da · CS intern-ready
↗
github · vkkatariya
contributions · 2026 51
skills
Python
90
ML / AI
82
TypeScript
68
Docker
76
Linux
80
SvelteKit
55
6 core · always building
now
orlon-bot
QLoRA · training
↗
homelab
2/3
nodes online
↗
stack
SvelteKitFE
FastifyAPI
PythonML
DockerOPS
TailscaleNET
Radxa 5THW
SQLiteDB
featured project
01 / 04
Finance
Buddy
Personal finance tracker built from 993 real transactions across 44 months. Spending analysis, category breakdown, waste audit, savings scenarios, and live search.
SvelteKit Fastify Postgres Drizzle Chart.js
view → github
projects
8
shipped
featured · hermes oauth
HERMES ONE
OAUTH FORK
13
PRS MERGED
4
PHASES
38/38
TESTS PASS
VIEW → GITHUB →
about

I'm Vishal — a CS student at Hochschule Darmstadt, building production systems while studying.

I work across the full stack: fine-tuning language models, building REST and WebSocket backends, designing interfaces, and managing self-hosted services on real ARM hardware. The common thread is building things that actually run — not just demos.

Currently based in Darmstadt. Looking for internships where I can work on systems that matter — flexible across ML engineering, full-stack development, and DevOps/infrastructure.

contact
github
vkkatariya
email
vishalkatariya404@gmail.com
📍 Darmstadt
open to werkstudent jobs
↓ download resume

timeline

Oct 2024education
Started CS at h_da
Hochschule Darmstadt · CS · since Oct 2024
2024homelab
First homelab — Rock 5T
Radxa Rock 5T · RK3588J ARM64 · 16 GB LPDDR5
Docker · Tailscale · self-hosted from scratch
2025project
Finance Buddy
993 real transactions · 44 months · Oct 2022–May 2026
SvelteKit + Fastify + Postgres + Drizzle + Chart.js
2025homelab
Homelab v2 — Proxmox cluster
+ Dell OptiPlex 7060 · Proxmox VE · 2-node cluster
KVM / libvirt · LXC containers · 48 GB total RAM
2026collab
TypeShift
Cross-platform AI writing assistant · with nayalambaliya
Android (Kotlin) · macOS (Swift) · Windows (C#)
2026AI · ML
orlon-bot — in progress
Telegram bot · QLoRA fine-tune on Kaggle T4 GPUs
Unsloth · inference on Rock 5T NPU (6 TOPs)
2028goal
B.Sc. Computer Science
Hochschule Darmstadt · expected graduation

all projects

view all →
02 / 04
Homelab Dashboard
Realtime ops dashboard for a 2-node homelab cluster. Live CPU/RAM/network metrics via WebSocket, NothingOS design, Notion docs viewer and service registry.
FastifyWebSocketDockerTailscale
App · private
03 / 04
TypeShift
Cross-platform AI writing assistant built in collaboration with nayalambaliya. Android (Kotlin), macOS (Swift), Windows (C#) — shared AI backend for text transformation.
KotlinSwiftC#AI
Collab · open source
04 / 04
orlon-bot
Telegram bot with a custom QLoRA fine-tuned model. Trained on Kaggle T4 GPUs via Unsloth. Inference on Rock 5T NPU (6 TOPs). Personal AI assistant — in progress.
PythonQLoRAUnslothTelegram
AI · ML · in progress
vkkatariya · work

Projects

Things I've built and shipped — from personal finance trackers to homelab infrastructure to collaborative AI tools.

01 / 08
Finance Buddy
Personal finance tracker
SvelteKit PostgreSQL Chart.js
Prototype complete
↗
02 / 08
Homelab Dashboard
Realtime ops · 2-node cluster
Fastify WebSocket Docker
Prototype complete
↗
03 / 08
TypeShift
Cross-platform AI · collab
Kotlin Swift C#
In progress
↗
04 / 08
orlon-bot
QLoRA fine-tune · Telegram
Python QLoRA Unsloth
In progress
↗
05 / 08
Portfolio Website
This site · Vercel + Tailscale
SvelteKit Vercel Tailscale Liquid Glass
Live
↗
06 / 08
Hermes One (OAuth fork)
Hermes One · OAuth for gated dashboards
Electron TypeScript React 18 Vite Tailwind better-sqlite3
13 PRs merged
↗
07 / 08
OpenClaw Dashboard
Control UI for OpenClaw gateway
React TypeScript WebSocket Neopop
Active dev
↗
08 / 08
UNILOX Fitness AI
AI Gym & Fitness Assistant · 7 modules
Next.js 14 FastAPI MediaPipe HuggingFace PostgreSQL MQTT
Phase 0 spec
↗
01 / 04

Finance Buddy

Personal finance tracker built from real transaction data — 993 entries across 44 months of actual spending and income.

Prototype complete Private · Tailscale only 2026
993
transactions
44
months of data
€43.7k
total tracked
7
dashboard tabs
spending breakdown · Oct 2022 – May 2026
The problem

I had 44 months of real transaction data exported from the Buddy app — but no way to see patterns across it. Which categories were eating my budget? What was genuinely avoidable? When would I realistically save €10k?

Spreadsheets didn't cut it. I needed a dashboard built specifically around my data.

What I built

A 7-tab self-hosted dashboard that parses, stores, and visualises all 993 transactions. Each tab answers a different question: where does the money go, is my spending trending up, what can I cut, and when will I hit my savings goals.

The Waste Audit tab alone surfaced €3,656 in avoidable spending I hadn't quantified before. The Savings Plan tab runs 3 scenarios — status quo, budget mode, and aggressive — showing projected savings dates.

7 tabs · what each one does
📊
Overview
KPI sparklines · monthly flow bar+line chart · category donut · My Tracker widget with live add-expense form
📈
Trends
44-month spending flow · year-by-year grouped bars · stacked category composition chart
🎯
Budget
13 categories · avg monthly bars · over/under visual for each
💼
Income
7 income sources breakdown · employer cards · earned vs passive yearly chart
🗑️
Waste Audit
€3,656 avoidable spend identified · category bars · 6 corner-cutting guide cards
💰
Savings Plan
3 scenarios (status quo 2047 · budget 2031 · aggressive 2028) · 84-month trajectory chart
🔍
Transactions
44 months navigable · day-grouped Buddy-style list · arc doughnut · live debounced search across all 993 entries
stack
SvelteKit + TypeScript
90
Fastify (REST API)
88
PostgreSQL 16
85
Chart.js 4.x
75
Docker + Caddy
70
access + privacy

Finance Buddy is private by design — it contains real financial data and is only accessible through Tailscale on authorised devices. The route buddy.auxois-wyrm.ts.net never touches the public internet. No public ports, all traffic through WireGuard.

02 / 04

Homelab Dashboard

Realtime ops dashboard for a self-hosted 2-node cluster — live metrics, service registry, and a Notion docs viewer.

Prototype complete Private · Tailscale only 2026
2
physical nodes
48 GB
total RAM
6+
services running
3
dashboard tabs
infrastructure · 2-node cluster
athena
Radxa Rock 5T · RK3588J ARM64 · 16 GB LPDDR5 · 6T NPU
OpenWebUI :3000 OpenClaw :18789 Stirling PDF :8080 Cockpit :9090 Home Assistant (KVM) Ollama :11434
Tailscale
WireGuard
atlas
Dell OptiPlex 7060 · i7-8700 x86_64 · 24 GB DDR4 · Proxmox VE
Proxmox WebUI :8006 Prometheus LXC ⬡ Grafana LXC ⬡ Gitea LXC ⬡ k3s VM ⬡ planned
Why I built it

Once you're running 6+ services across 2 nodes, `ssh + docker ps` stops scaling. I wanted a single pane of glass for the entire cluster — live metrics, service status, and the homelab docs I write in Notion, all in one place.

The NothingOS design system (dark `#1B1C1D`, `#E2201F` red, dot-matrix background) made the dashboard feel like a proper product rather than a dev tool.

What's in each tab
⚡
Infrastructure
CPU pie meters · RAM bar · storage verticals · network sparkline · services grid · uptime rings
🔗
Services
Service registry with URLs, ports, node tags · filter by node · access URL list
📄
Docs & Blueprints
Notion homelab/DIY pages rendered as cards · slide-in detail panel with tabbed sections
stack
Fastify + TypeScript
90
WebSocket (ws)
85
Tailscale file-cert
85
Docker + Caddy
80
NothingOS CSS
70
03 / 04

TypeShift

Cross-platform AI writing assistant — same AI, three native apps. Built in collaboration with nayalambaliya.

In progress Open source Collaboration
3
platforms
3
languages
2
contributors
1
shared AI backend
three platforms · one experience
🤖
Android
Kotlin · Jetpack Compose
Native Material You design · full keyboard access · background AI processing
🍎
macOS
Swift · SwiftUI
Native Mac experience · menu bar integration · system-wide text access
🪟
Windows
C# · WinUI 3
Fluent Design · context menu integration · clipboard-aware transforms
The idea

Writing assistants are either web-only or locked to one OS. TypeShift is natively built for each platform — not a web wrapper — with a shared AI backend powering transformations across all three.

Select text anywhere on your device, invoke TypeShift, choose a transform (rewrite, summarise, improve tone), and get the result back in place.

My role

Joined as a collaborator on the project led by nayalambaliya. Contributed to the macOS CI release workflow, project scaffolding, and dev setup documentation. The cross-platform architecture decision — shared API backend, native UI per platform — was established early and I've helped build toward it.

Repo: github.com/nayalambaliya/TypeShift

stack
Swift / SwiftUI
85
Kotlin
80
GitHub Actions
75
Shared AI API backend
70
C# / WinUI 3
65
04 / 04

orlon-bot

A Telegram bot with a custom fine-tuned LLM — trained on Kaggle T4 GPUs using QLoRA and Unsloth, running inference on a Rock 5T NPU.

In progress ML / AI 2026
30h
Kaggle T4 / week
6 TOPs
Rock 5T NPU
QLoRA
fine-tune method
Unsloth
training framework
ML pipeline · train to inference
01
Dataset
Custom training data · curated for personal assistant behaviour
›
02
Kaggle T4
Free 30h/week T4 GPU · Jupyter notebook environment
›
03
Unsloth
2× faster training · 70% less VRAM · QLoRA adapters
›
04
Model export
GGUF format · quantised for edge deployment
›
05
Rock 5T NPU
RK3588J · 6T NPU · rkllama inference on athena
›
06
Telegram
Bot API · Python telegram-bot library · always-on via pm2
Why QLoRA + Unsloth

Full fine-tuning a 7B model requires 80 GB+ of VRAM — not available on free Kaggle T4s (16 GB). QLoRA fine-tunes only a small set of low-rank adapter matrices, cutting memory usage by ~4×. Unsloth adds another 2× speed improvement through kernel-level optimisations, making the Kaggle free tier genuinely viable.

The result: a model that responds in my style, knows my context, and runs entirely on hardware I own.

The NPU angle

The Rock 5T has a Rockchip RK3588J NPU with 6 TOPs — enough for inference on quantised models in the 1–7B range. Using rkllama, the model runs natively on the NPU rather than the CPU, giving better performance per watt.

The whole pipeline — training on Kaggle, deploying on the NPU, serving through Telegram — closes a loop between free cloud compute and self-hosted inference.

stack
Telegram Bot API
95
Python
90
Unsloth
85
QLoRA (PEFT)
85
Kaggle T4 GPU
80
05 / 08

Portfolio Website

This site · Vercel + Tailscale

Live Hybrid · Vercel + Tailscale 2026
7
pages · single SPA
3
design systems
8
project cards on top
2
deploy targets
deploy topology · vercel + athena
vercel
vishalkatariya.dev · global CDN edge network
/ · home /projects · /roadmap /about /me → redirect to gate
● public · open
request
routing
athena
Radxa Rock 5T · Caddy + Tailscale IP allowlist
/me/* · gated dashboard auxois-wyrm.ts.net no public ports
● private · tailscale only
Hybrid Architecture

Personal portfolio at vishalkatariya.dev — built as a self-contained SvelteKit SPA, hybrid-hosted with public routes on Vercel and private /me/* behind a Tailscale + Caddy IP allowlist on athena.

Design system uses three layered surfaces: liquid glass (nav/frames), neomorphism (widgets), and NeoPOP (CTAs). Typography stacks Cormorant Garamond italic (initials) with Space Grotesk (display), Outfit (body), DM Mono (labels), and NDOT (accent).

Consolidated SPA

Every page is a single HTML prototype in prototypes/ — portfolio-combined.html is the consolidated SPA, with separate projects.html, about.html, and cs-roadmap.html still in use as standalone entry points.

Features
📐
Hybrid Hosting
Public routes on Vercel CDN + private /me/* behind Tailscale + Caddy IP allowlist on athena
🎨
3 Design Systems
Liquid Glass (nav/frames), Neomorphism (widgets), NeoPOP (CTAs) layered together
⚡
Single SPA
portfolio-combined.html consolidates all 5 standalone prototype pages into one experience
stack
Tailscale
95
Vercel
85
SvelteKit
85
Liquid Glass
75
06 / 08

Hermes One OAuth Fork

Hermes One · OAuth for gated dashboards

13 PRs merged Open source · fork Phase 4 pending 2026
13
PRs merged
4
phases shipped
38
tests pass
1
upstream PR pending
OAuth fork · 4 phases · 13 PRs
phase 01
Audit
Codebase review · dependency map · ?token= auth gap identified
✓ 3 PRs merged
›
phase 02
OAuth Flow
Nous Portal OAuth · /api/status provider discovery · main-process cookies
✓ 5 PRs merged
›
phase 03
Mac E2E
Single-use WS ticket minting · Vitest suite · 38/38 tests pass
✓ 5 PRs merged
›
phase 04
Upstream PR
fathah/hermes-desktop:main · community merge · cross-provider
⬡ pending
The Problem

A patch project on top of fathah/hermes-desktop — the community Electron desktop client for Hermes Agent — that adds the Nous Portal OAuth login flow plus single-use WebSocket ticket minting so the app can connect to a Hermes dashboard bound to 0.0.0.0 in gated mode.

Without this patch, the community client's ?token= URL-based auth can't reach a Nous Portal-gated dashboard: the dashboard returns 302 to the login page because the spawned backend is a separate Node process without the app's main-process OAuth cookies. The fix is in src/main/oauth.ts and src/main/dashboard.ts — provider discovery via /api/status + ticket-based WS auth that doesn't burn the single-use ticket before the renderer can use it.

What I Built

13 PRs shipped across 4 phases: Phase 1 audit, Phase 2 OAuth dashboard auth flow, Phase 3 Mac end-to-end, and Phase 4 (pending) upstream PR to fathah/hermes-desktop:main. 38/38 OAuth tests pass.

view on GitHub →

Features
🔐
Nous Portal OAuth
Replaces ?token= URL auth with proper OAuth code path so the community app can talk to gated dashboards
🎟️
Single-Use WS Tickets
Mints per-connection WS auth tickets that don't get burned by probes before the renderer can use them
🔍
Provider Discovery
Discovers dashboard auth providers from /api/status instead of hardcoding nous — community-portable
🧪
38/38 Tests
Vitest suite covers IPC channels, OAuth flow, WS ticket lifecycle, dashboard connection
stack
Electron
95
React 18
85
Vite
80
Tailwind
75
better-sqlite3
70
07 / 08

OpenClaw Dashboard

Control UI for OpenClaw gateway

Active dev Private · homelab 2026
3
routes
7
agents orchestrated
2
themes
1
gateway port
3 views · WebSocket RPC gateway
/dashboard
channels · instances · sessions
Agent status grid · metrics overview · active session counter · live via WS RPC
/chat
slash commands · 7 agents
/network · /agent slash commands · conversational interface across orchestrated agents
gateway
ws://127.0.0.1:18789
Ed25519 device identity · WS RPC · OpenClaw orchestration layer on athena
Overview

React + TypeScript control UI for the OpenClaw gateway — the local-first agent orchestration layer running on athena. Connects to the gateway over WebSocket RPC with Ed25519 device identity stored in localStorage.

AppShell layout with Sidebar + Topbar; routes for /dashboard (channels, instances, sessions metrics), /chat (with slash commands like /network and /agent), and /theme-test.

Design System

Hybrid NeoPOP + Neumorphism design system with two themes (hybrid-dark primary, hybrid-light). Theme tokens in src/theme/tokens.ts. Command palette (src/components/commands/) for keyboard-first navigation across the registered agents.

Features
🌐
WebSocket RPC
Ed25519 device identity + RPC client connecting to OpenClaw gateway over ws://127.0.0.1:18789
🎨
NeoPOP + Neumorphism
Hybrid design system via @cred/neopop-web, tokens in src/theme/tokens.ts, two themes (dark + light)
⚡
Command Palette
Keyboard-first navigation in src/components/commands/, slash commands like /network and /agent in chat
stack
React
95
TypeScript
90
WebSocket
85
Neopop
75
08 / 08

UNILOX Fitness AI

AI Gym & Fitness Assistant · 7 modules

Phase 0 spec Private · commercial AI · ML 2026
7
AI modules
3
deploy targets
5
IoT edge devices
0
shipping yet
architecture · edge → backend → AI modules
edge
IoT devices · MQTT (Mosquitto) · Node-RED
Jetson Nano · pose CV Raspberry Pi 5 · AI inference ESP32-CAM · video stream
MQTT
pub/sub
backend
FastAPI · MLflow · HuggingFace · Caddy
FastAPI :8000 MLflow :5000 MediaPipe · Transformers
REST
WebSocket
AI modules · 7
Next.js 14 · React Native (Expo)
Gym Trainer (CV · MediaPipe) Dietician (NLP · diet plans) Habit Tracker (behavioral ML) Virtual Gym Buddy (conv. AI) + 3 more ⬡ planned
Overview

Unified AI-powered fitness ecosystem by Revolux Learning Private Limited. Seven core modules: AI Gym Trainer (CV pose estimation + rep counting + form correction), AI Dietician (NLP diet recommendations + meal planning), Smart Gym Assistant (IoT equipment integration via MQTT + Node-RED), AI Habit Tracker (behavioral ML predicting skipped workouts), Virtual Gym Buddy (conversational AI + sentiment analysis), Pose-to-Performance Analyzer (motion efficiency + biomechanical reports), and Gym Recommender.

Phase 0 (specification & design) is complete. Phase 1+ (foundation scaffold, core AI modules, integration) not started — staged via monorepo with apps/web, apps/mobile, apps/api, ml/, iot/, infrastructure/, and shared/ workspaces.

Architecture

Stack spans Next.js 14 web dashboard + React Native (Expo) mobile + FastAPI backend, with Hugging Face transformers + MediaPipe for ML, MQTT (Mosquitto) for IoT edge devices (Jetson Nano / RPi 5 / ESP32-CAM), MLflow for experiment tracking, and Caddy reverse proxy in production.

Features
📷
Computer Vision
MediaPipe Pose + custom CNN heads for rep counting, form scoring, biomechanical analysis
🤖
7 AI Modules
Combines CV + NLP + IoT + behavioral ML + conversational AI in a single ecosystem
📡
Edge + Cloud
MQTT (Mosquitto) on Jetson Nano / RPi 5 / ESP32-CAM, FastAPI backend, MLflow for experiment tracking
stack
MQTT
85
Next.js 14
80
FastAPI
80
MediaPipe
75
PostgreSQL
70
← home ↑ back to top
vkkatariya · roadmap

Roadmap

Your Complete Guide to Mastering Computer Science Fundamentals

📋 11 Core Topics ⏱ 12–14 Months 🤖 ML · Infra · Full-stack
Start Your Journey → ← back to portfolio

learning path overview

01
Foundation
Months 0–3
Programming Fundamentals OOP OOD
02
Core Skills
Months 3–7
DS&A Design Patterns Software Engineering Databases
03
System Core
Months 8–10
OS Networking Computer Architecture
04
Security & Specialization
Months 11–14
Security Fundamentals Career Path

getting started

Where to Start

  1. Choose C++ as your first language
  2. Focus on concepts, not syntax
  3. Start with simple programs
  4. Practice daily — even 1 hour makes a difference
  5. Join a community

Tips for Success

  • Be patient — progress is non-linear
  • Type code, don't copy-paste
  • Read error messages carefully
  • Build projects from the start
  • Take breaks — rest is learning
  • Celebrate small wins

Common Mistakes to Avoid

  • Jumping between languages constantly
  • Tutorial hell — watch but never build
  • Comparing yourself to others
  • Skipping fundamentals for frameworks
  • Not asking for help
  • Giving up too early

month-by-month timeline

  • Variables & data types
  • Control structures (if/else, loops, switch)
  • Functions & scope
  • Arrays & strings
  • Pointers & references
  • Daily: 2–3 hrs minimum · 3–5 problems/day · mini project weekly
  • Classes & objects
  • Inheritance & polymorphism
  • Encapsulation & abstraction
  • Constructors & destructors
  • Virtual functions & abstract classes
  • Build 2–3 OOP projects (Bank, Library, Game)
  • SOLID principles (Single Responsibility, Open/Closed, Liskov, Interface Segregation, Dependency Inversion)
  • UML diagrams (class, use case, sequence, activity)
  • Design a complete system from scratch
  • Code reviews & design trade-off analysis
  • Daily: 1–2 hrs · diagram at least one system per week
  • Arrays, Linked Lists, Stacks, Queues
  • Trees (Binary Trees, BST, AVL, Heap)
  • Graphs (BFS, DFS, shortest paths)
  • Hash Tables & Hash Maps
  • Sorting & Searching algorithms
  • Dynamic Programming (Memoization, Tabulation)
  • Greedy algorithms · Big O Analysis (Time & Space)
  • Target: Solve 150+ LeetCode problems
  • Creational patterns: Singleton, Factory Method, Abstract Factory, Builder, Prototype
  • Structural patterns: Adapter, Decorator, Facade, Proxy, Composite
  • Behavioral patterns: Observer, Strategy, Command, Iterator, Template Method
  • 23 classic Gang of Four (GoF) patterns
  • Apply patterns to real projects · Refactor existing code
  • Clean code principles & refactoring techniques
  • Git workflows (branching, PRs, code reviews)
  • Testing (unit, integration, TDD)
  • Documentation best practices
  • Agile/Scrum & Kanban methodologies
  • DevOps & CI/CD basics · Contribute to open source
  • SQL fundamentals (DDL, DML, JOINs, aggregate functions)
  • DB design & normalization (1NF, 2NF, 3NF)
  • ER diagrams, primary & foreign keys
  • Indexes, views & performance optimization
  • NoSQL basics: MongoDB, document model
  • SQL vs NoSQL trade-offs · Redis basics
  • CPU internals: ALU, Control Unit, Registers
  • Instruction Fetch-Decode-Execute cycle
  • Memory hierarchy: Cache L1/L2/L3, RAM, virtual memory
  • Pipelining & branch prediction
  • Performance optimization & cache-friendly code
  • Modern architectures: RISC vs CISC, multi-core, GPU basics
  • Process & thread management
  • CPU scheduling algorithms
  • Memory management & virtual memory
  • File systems & I/O management
  • Deadlocks, synchronization (mutex, semaphore)
  • Linux basics: command line, shell scripting, process management
  • OSI model: all 7 layers explained
  • TCP/IP suite · UDP vs TCP
  • HTTP/HTTPS, DNS, DHCP, TLS/SSL
  • IP addressing & subnetting
  • Network security fundamentals
  • Build a networked application · Wireshark packet analysis
  • Cryptography: hashing (SHA-256, bcrypt), encryption (AES, RSA)
  • OWASP Top 10: injection, XSS, broken auth, broken access control
  • Authentication & authorization (JWT, OAuth)
  • Secure coding practices & input validation
  • API security & HTTPS everywhere
  • Practice on DVWA or WebGoat · security audit your own projects
  • Choose your career path (Frontend, Backend, ML, DevOps, etc.)
  • Build a portfolio of 3–5 substantial projects
  • Interview preparation: DS&A, system design, behavioral
  • Open source contributions
  • Start applying to internships and entry-level roles
  • Network, attend meetups, build your online presence

11 core topics

career paths

curated resources

Programming & OOP

C++ Primer 5th Edition
Beginner–Intermediate ↗
Programming: Principles and Practice Using C++ (Stroustrup)
Beginner ↗
OOP in C++ (Lafore)
Intermediate ↗

Data Structures & Algorithms

Introduction to Algorithms — CLRS
Advanced ↗
Cracking the Coding Interview (McDowell)
Intermediate ↗
Algorithm Design Manual (Skiena)
Intermediate ↗

System Design & Architecture

Clean Architecture (Robert C. Martin)
Advanced ↗
Design Patterns — GoF
Advanced ↗
Head First Design Patterns (O'Reilly)
Intermediate ↗
Progress
0 / 11
vkkatariya · personal

About

Vishal Katariya
open to werkstudent jobs
Download Resume
PDF · A4 · 2026
↓

I'm Vishal Katariya — a CS student at Hochschule Darmstadt, based in Darmstadt. I build across the full stack: fine-tuning language models, wiring REST and WebSocket APIs, and managing self-hosted infrastructure on real hardware.

My projects run in production — not just on localhost. From a personal finance tracker built on 993 real transactions, to a 2-node Proxmox cluster running 6+ services, to a Telegram bot trained with QLoRA on a free Kaggle GPU and deployed to an ARM NPU.

The common thread: I care more about how systems actually work than what they look like on the surface. Open to internships in Germany across ML engineering, full-stack development, and infrastructure roles.

Education
01
HOCHSCHULE DARMSTADT · DARMSTADT, GERMANY
B.Sc. Computer Science
Oct 2024 – Expected 2028
Informatik Darmstadt Germany
currently enrolled
MODULES COMPLETED / IN PROGRESS
Algorithms & Data Structures Object-Oriented Programming Mathematics for CS Computer Architecture Database Systems Operating Systems Software Engineering Computer Networks
Skills
02
core technical
Python
86
Docker · Linux
82
TypeScript
68
ML / AI
72
SvelteKit
60
Proxmox · ARM
64
C++
52
Backend & Infrastructure
Fastify WebSocket REST API PostgreSQL Drizzle ORM SQLite Docker Compose Caddy Tailscale Proxmox VE KVM / libvirt ARM64 pm2 UFW
AI & Machine Learning
QLoRA fine-tuning Unsloth Ollama LLM inference rkllama (NPU) Hugging Face Kaggle T4 GPU PyTorch GGUF / quantisation Transformers
Frontend
SvelteKit TypeScript Chart.js Vanilla CSS Motion One Responsive design Liquid Glass CSS NothingOS design React (learning)
Soft Skills
Problem-solving Research & analysis Documentation Teamwork Self-directed learning Async collaboration Technical writing Systems thinking
Languages
03
🇩🇪
Deutsch
PROFESSIONAL WORKING PROFICIENCY · B2–C1
Working language in Germany · handle professional conversations, emails, and technical discussions · studying at a German-language institution
🇬🇧
English
FULL PROFESSIONAL PROFICIENCY · C1–C2
Primary language for all technical work, documentation, code, and AI tooling · full professional fluency
🇮🇳
Hindi
FULL PROFESSIONAL PROFICIENCY · NATIVE
Native speaker · full fluency across formal and informal contexts
🇮🇳
Gujarati
FULL PROFESSIONAL PROFICIENCY · NATIVE
Native speaker from Surat, Gujarat · full fluency including regional dialect
Interests
04
💻
Programming
Started coding in 10th grade — before I knew it was a career path. Now building production systems, self-hosted infrastructure, and ML pipelines across whatever stack fits the problem.
🤖
Artificial Intelligence
Not just using AI tools — understanding how they work under the hood. Fine-tuning LLMs with QLoRA, running inference on edge NPUs, building multi-agent systems on self-hosted hardware.
🏏
Cricket
Played since childhood — always been a big part of life. The strategy, the patience, the team dynamics. Grew up watching and playing in India and still follow it closely.
📚
Entrepreneurship
Reading business books and studying how companies are built — not just the tech, but the strategy, the product thinking, and what makes a system scale. Building toward something called orlon.
VK

Private section

/me is only available on the Tailnet.
Connect to Tailscale and open the private site.

me.auxois-wyrm.ts.net