bash — 80x24 $ python train.py --epochs 10 Loading Phi-4 14B weights... Epoch 1/10 loss=0.4821 Epoch 2/10 loss=0.3104 Epoch 3/10 loss=0.2291 Epoch 4/10 loss=0.1887 Epoch 5/10 loss=0.1432 Validation R2 = 0.9650
SIDE A 45 RPM

AI Developer  ·  Melbourne

MATHEESHA
ILEPERUMA

Training models, not opinions.

See my work ↓
Scroll

Building
at the
frontier.

I'm Matheesha Ileperuma, a final-year Computer Science (AI) student at Swinburne University of Technology, Melbourne. I design and build AI systems that actually ship — offline LLMs for remote communities, production-grade recommendation engines with full CI/CD pipelines, and ML classifiers deployed behind real APIs.

My work spans RAG pipelines, fine-tuned language models, full-stack web applications, and everything in between. When I'm not training models, I'm chasing down a JDM part or digging through a record crate.

Currently looking for grad roles and internships in AI/ML engineering — Melbourne-based or remote.

DegreeB.Comp.Sc. (AI), Swinburne University
Graduating2025
BasedMelbourne, Australia
Open toGrad Roles · Internships · Remote

Projects.

02 / 06

Capstone · University

TEK Climate System

Offline LLM chatbot preserving traditional ecological knowledge for an indigenous community in Bangladesh. Runs entirely on local hardware — zero internet dependency. Led RAG pipeline design and LLM evaluation (Mistral 7B vs Phi-4 14B); wrote sprint documentation and ran client demos.

Phi-4 14B Ollama RAG ChromaDB all-MiniLM-L6-v2 Flask Docker SQLite

03 / 06

ShopWise AI

Content-based product recommender across 125 products in 14 categories. Builds a TF-IDF matrix from combined product features, computes cosine similarity, and explains each suggestion in plain English. No framework — pure JS frontend.

Python Flask scikit-learn TF-IDF Cosine Similarity Vanilla JS
GitHub →

04 / 06

Personal Finance Categoriser

Upload a CSV of bank transactions → ML classifies spending categories → interactive Recharts dashboard. User corrections feed back into retraining. Trained on 1,600 labelled transactions, 100% test accuracy.

React 18 Vite FastAPI Naive Bayes TF-IDF Recharts SQLite

05 / 06

Stock Price Prediction

AAPL forecasting with LSTM, GRU, SimpleRNN, and an ARIMA ensemble built across 7 checkpoints of increasing complexity. Best model: Bidirectional GRU (2 layers, 128 units). Baseline LSTM R² = 0.965. Includes 5-day multistep forecasting.

TensorFlow / Keras LSTM · GRU ARIMA statsmodels yfinance scikit-learn

06 / 06

AI4Cyber

ML applied to cybersecurity: spam detection (TF-IDF + Logistic Regression / Naive Bayes / Linear SVM) and malware detection (PE header features + Random Forest). Full EDA reports with word clouds, class distributions, and cross-model comparisons.

Python scikit-learn Random Forest SVM WordCloud matplotlib joblib

Skills.

AI / LLM

PyTorch
TensorFlow
LangChain
Hugging Face
Ollama
Anthropic / Claude API
RAG Pipelines
pgvector · ChromaDB

Machine Learning

scikit-learn
LSTM · GRU · RNN
ARIMA / Time Series
TF-IDF · Cosine Sim
Naive Bayes · SVM
Random Forest
pandas · NumPy

Backend

Python
FastAPI · Uvicorn
Flask
PostgreSQL 16
SQLite · SQLModel
JWT Auth

Frontend

React 18
TypeScript
Vite
Tailwind CSS
Recharts
Vanilla JavaScript

Tools & Infra

Git · GitHub
Docker
GitHub Actions CI/CD
Linux
Vercel · Render

Let's build
something.

Open to grad roles, internships, and interesting problems — Melbourne or remote.