MSc Data Science · King's College London · 2026

Arvinth
Srinivasasekar

Fine-tuning transformers. Shipping production ML. Building AI that explains itself.

0 Training records
0 Attack classes
0 Classifier accuracy
0 Real-time recognition
DistilBERT Attention Weights · Layer 6 · Head 4
IoT Intrusion Detection · 720,927 network flows · 34 attack classes

LLM Fine-Tuning for
IoT Intrusion Detection

DistilBERT fine-tuned on 720,927 network flow records — a novel parser-free approach that treats tabular security data as natural language, with SHAP-based interpretability to explain every prediction.

SHAP Feature Importance
Top drivers of attack classification predictions
Protocol type
0.847
Duration
0.723
Source bytes
0.691
Service
0.634
TCP flag
0.612
Destination bytes
0.578
Land flag
0.541
Wrong fragments
0.498
81.97%
Accuracy
34-class imbalanced classification
63.33%
Macro F1
SMOTE on rare attack classes
720K+
Records
IoT network flows processed
Top 20
SHAP features
Identified via explainability layer
Stack
PyTorch HuggingFace DistilBERT SHAP XGBoost SMOTE Scikit-learn

Things I've built

2026 Production ML

Autonomous AI Stock Trading Engine

Weighted soft-vote ensemble (XGBoost, LightGBM, Logistic Regression) evaluating Indian equities across 34 technical and fundamental features. Includes a Market Regime Classifier to reduce drawdown risk, SHAP white-box explainability on every prediction, and a real-time VADER NLP sentiment pipeline via n8n. Full inference stack exposed through FastAPI, with PostgreSQL for backtest storage and Docker for portability.

Market Data
Feature Eng.
XGB / LGB
SHAP
FastAPI
Signals
News Feed
VADER NLP
n8n
PythonXGBoostLightGBMFastAPIPostgreSQLDockern8nSHAP
Dec 2023 – Apr 2024 Computer Vision · NLP

Gesture & Voice Recognition System

Led a 4-member team to build a real-time gesture and voice recognition control system for accessibility applications. Combined computer vision (OpenCV) and NLP-based speech recognition to handle 50+ unique commands. Dataset augmentation, noise reduction, and hyperparameter tuning improved model responsiveness by 25%.

92% Real-time gesture accuracy
+25% Response via augmentation
50+ Voice commands processed
PythonTensorFlowOpenCVNLPSpeech Recognition
2026 Agentic AI · Automation

Agentic Job Hunter

Built for my own job search because manually tracking applications was absurd. Runs at 9am daily via macOS launchd: Apify LinkedIn scraper + 50-employer career page scraper (Greenhouse, Lever, Ashby, Workday) → ATS keyword extraction from job descriptions → UK sponsor register check across 125,000 entries → DDG recruiter enrichment → openpyxl workbook with conditional formatting. Zero cost, no LLM required for the daily run.

50+ Company pages scraped daily
125K UK sponsor register entries
£0 Daily running cost
PythonBeautifulSoupApifyopenpyxlmacOS launchd
Aug – Dec 2024 Healthcare ML

Diabetes Risk ANN

ANN achieving 87% prediction accuracy on 10,000+ patient records. Feature selection and normalisation pipelines improved over baseline by 15%.

TensorFlowScikit-learn
Dec 2024 – Mar 2025 Time Series

Solar Forecasting Ensemble

Stacking ensemble (GRNN, ENN, BPN) for solar power generation. 95% accuracy over 3 years of data; 18% improvement over single-model baselines.

PythonMATLAB
Nov 2024 Blockchain

Blockchain Voting System

Tamper-proof election prototype on Ethereum testnet. 20% transaction time reduction via smart contract optimisation. Presented at Tamil Nadu Government Academic Conference.

EthereumSmart Contracts

What I work with

ML & AI
PyTorch HuggingFace Transformers XGBoost LightGBM TensorFlow Scikit-learn SHAP LangChain RAG OpenAI API
Engineering
Python FastAPI Docker PostgreSQL REST APIs n8n MCP Git Postman
Cloud & Data
Azure Azure ML AWS GCP Apache Spark SQL Power BI Pandas & NumPy

The journey

2025–2026
King's College London
Education

MSc Data Science

Advanced machine learning, big data analytics, statistical modelling. Dissertation: LLM fine-tuning for large-scale IoT intrusion detection. London, UK.

Nov–Dec 2023
Arus Info Pvt Ltd
Experience

Data Analytics Intern · Bangalore

Automated 5+ reporting workflows in Power BI and Excel, cutting processing time by 30%. VBA and Python scripts reduced data entry by 40% across a 5-member analytics team.

2021–2025
VIT Vellore
Education

B.Tech Computer Science & Engineering (IoT)

CGPA: 8.35/10.0. Foundation in AI/ML, neural networks, data systems, and IoT. Best Presenter Award at NTU Singapore, 2023.

National University of Singapore
AI-Powered Business Analytics
Jul 2023
Nanyang Technological University
Introduction to Artificial Intelligence
Mar 2023
IBM
Data Analytics Externship
Dec 2022

Let's talk

I'm looking for London-based AI/ML/Data roles starting mid-2026. If you have a role or just want to connect, reach out directly.

Visa Status
UK Graduate Route
from September 2026
2 years full-time work rights · No sponsorship required · No cost or paperwork for the employer
Currently: Student visa · MSc Data Science · King's College London