Vighnesh Sadvilkar

MSc Adaptive Cybersecurity, University of Galway

Galway, Ireland · Open to relocating within Ireland
Email me Download CV GitHub LinkedIn

Profile

I'm finishing an MSc in Adaptive Cybersecurity at the University of Galway, with a B.Eng. in Computer Engineering from the University of Mumbai before that. Most of my project work sits at the intersection of security and machine learning: training classifiers, stress-testing them against adversarial input, and figuring out why they fail before someone else finds out for me. I'm comfortable across the stack, from PyTorch model training to REST APIs to Linux and packet-level tools like Wireshark and Suricata.

Education

University of Galway2025 – 2026

MSc Computer Science: Adaptive Cybersecurity, Galway, Ireland

  • Modules: Network Security, Cryptography, Threat Analysis, Secure Software Design
  • Applied coursework in data-driven systems, model evaluation, and secure application design
University of Mumbai2021 – 2025

B.Eng. Computer Engineering, Mumbai, India

  • Coursework in software engineering, data structures, algorithms, and systems programming
  • Final-year project: a decentralised secure communication platform built on blockchain

Technical Skills

Programming
Python, Java, JavaScript, C, SQL
ML & Data
PyTorch, Scikit-learn, Pandas, NumPy, Tableau
Web & APIs
REST APIs, JSON, HTML, CSS, Django
Databases
MySQL, PostgreSQL (basic), MySQL Workbench, SQLite
Security & Networking
Wireshark, Suricata, Snort, Splunk (basics), John the Ripper, TCP/IP fundamentals, Network Security, Cryptography, Threat Analysis
Tools
Git, VS Code, Linux, CLI, CI/CD basics

Selected Work

LLM Prompt Injection Detection System

PyTorch · HuggingFace Transformers · scikit-learn · 2026

A model-agnostic proxy layer that intercepts and classifies prompts before LLM ingestion, deciding Allow or Block. Benchmarked three model families on 20,662 prompts, reaching F1 0.947 and ROC-AUC 0.985 at 20ms per query.1 Found and fixed systematic label noise affecting the benchmark dataset.2

Evasion attack success rate, before vs. after input normalisation3
Before
100%
After
1–3%
False negatives on indirect injection, before vs. after fine-tuning4
Before
112
After
78
View repository →

Transfer Learning-Based Intrusion Detection System

PyTorch · Python · 2025

Trained a model reaching 98%+ accuracy classifying network attack types across two CIC-IDS datasets, with a cross-dataset pipeline studying how well results generalise from one year's traffic to the next.

View repository →

Secure Communication Platform (Blockchain)

Blockchain · Smart Contracts · Full-stack · 2024

B.Eng. final-year project: a full-stack decentralised messaging platform using cryptographic hashing to guarantee message integrity, with no single point of failure. It covers protocol design, smart contract logic, and a lightweight web interface, end to end.

Email Phishing Detection

Python · scikit-learn · 2024

A classifier trained on linguistic and structural features of phishing emails, reaching 92% accuracy, with an automated test pipeline measuring performance across datasets.

Notes & Password Manager

JavaScript · SQL · REST APIs · 2023

A cross-platform app with end-to-end encrypted storage and two-factor authentication, built with JavaScript on the frontend and SQL for persistence, talking to a REST API.

Certifications

Completed
  • Google Cybersecurity Professional Certificate
  • Google Data Analytics Professional Certificate
  • Cisco Networking Academy: Introduction to Cybersecurity
In progress
  • CompTIA Security+
  • Cisco Networking Academy: Networking Basics

Leadership & Achievements

Notes

  1. Best of three benchmarked model families (TF-IDF + LinearSVC, BiLSTM + GloVe, ModernBERT); ModernBERT-base was the strongest performer.
  2. Systematic label noise affected roughly 62% of benign-labelled rows in a public benchmark dataset; correcting it improved traditional ML F1 by roughly 12 points.
  3. Measured against homoglyph substitution, Base64 encoding, and Unicode Tag-block smuggling.
  4. A detector trained only on direct-injection examples saw ROC-AUC drop from 0.985 to 0.613 on indirect injection; warm-start fine-tuning closed most of that gap.