Browse Projects

I built a conversational AI assistant directly into my portfolio site. Recruiters and visitors can ask anything about my projects, skills, and background — and get grounded, cited answers in real time, powered by RAG (Retrieval-Augmented Generation) and Llama 3.3 70B via Groq.

Generates anime faces using a DCGAN trained on 63,000 images. Every face is produced from random noise — none of them exist. Trained on Google Colab T4 GPU. Includes latent space interpolation showing smooth morphing between faces.

1970s Boston house price predictor. Linear regression on 13 features with log-transformed target (test r²=0.74). Live slider UI on Azure; full analysis in rendered notebook.

Semantic CV vs job description analysis identifying missing skills and producingactionable gap recommendations. HuggingFace embeddings for match scoring, Pineconefor a searchable job library, ROUGE for ATS keyword alignment. Accepts jobdescriptions as text, PDF, or URL via trafilatura extraction.

Credit risk scoring API on the German Credit dataset. Three classifiers benchmarked; Logistic Regression wins by cost-weighted score. FastAPI serves predictions with plain-language explanations. Deployed on Azure via GitHub Actions CI/CD.

Draw a digit on the canvas — the CNN predicts it in real time. Trained on 60,000 MNIST images. Confidence score and probability bar chart across all 10 digits. Live on Azure App Service.

T5-small fine-tuned on the AESLC corpus (14k Enron email/subject pairs) to generate professional subject lines from an email body. Five suggestions returned at varying temperature levels — from beam-search conservative to top-p creative sampling.

Three economic signals — Tesla, Bitcoin, and US unemployment — each paired with Google Trends search data across mismatched time frequencies and aligned by resampling. The 2020 COVID shock compressed months of unemployment pattern into weeks.

Upload two CSV columns, select a test, get a plain-English verdict. Five hypothesis tests with automatic assumption checking. Datasets above 50,000 rows route through PySpark automatically.

LEGO nearly went bankrupt in 1998. This project joins six relational tables across 15,710 sets to measure exactly what changed — licensed share grew from 0% to 31.5%, average complexity 8×, and minifigure density 57% from 1970 to 2010.

PDF document summarizer using Amazon Bedrock (Claude Haiku) with production LLMOps patterns: versioned YAML prompt templates, automated ROUGE evaluation, Bedrock Guardrails for PII redaction, and MLflow tracking of every inference call. Deployed on Azure App Service.

Detects the language of any text across 22 languages using character n-gram TF-IDF features. Returns the top 3 languages with confidence scores and a plain-English explanation of the character patterns driving the detection.

Pipeline that ingests 180,896 real Hadoop cluster logs, detects error spikes using rate-based thresholding, and delivers results via a FastAPI dashboard, scheduled HTML email, and Slack alerts with state-machine deduplication.

Across 5,384 films from 1915–2018, a linear regression on budget vs. worldwide gross gives a slope of 3.12 (R²=55.77%) — every dollar of budget is associated with $3.12 in revenue, yet 37.28% of films still failed to recoup their costs.

Four-stage PySpark ETL pipeline on 40M+ NYC 311 service requests. Identical pandas pipeline for direct comparison. Benchmarked at 100k, 1M, and 10M rows — showing the exact crossover where Spark's parallelism overcomes its initialisation overhead.

NumPy treats images as numbers — a 768 × 1024 photograph is a 3D array of integers. This project works through ndarray operations from first principles: slicing, broadcasting, matrix multiplication, and pixel-level transforms like greyscale conversion and colour inversion.

Full-stack Flask portfolio with a custom admin dashboard, PostgreSQL, a REST API, an AI assistant powered by RAG, Cloudflare R2 media storage, and CI/CD deployment to Railway via GitHub Actions.

Re-examination of Dr Semmelweis's 1861 hospital records from Vienna General Hospital (1841–1849). Mandatory handwashing in June 1846 cut the average monthly death rate from 10.5% to 5.0% — confirmed statistically at p ≈ 0.00025.