MSc Computer Science

Overview

Portfolio Journal is an Android-oriented Flutter client built for the 7CC012 Mobile Application Development brief. It keeps a personal journal fully usable offline: notes live in SQLite on the device, images are copied into app documents, and sharing goes through the system share sheet rather than a custom social backend.

Capabilities

• Create, edit, view, and delete notes (optional title, required body)
• Home feed of messages with detail screens and search
• Swipe-to-delete and long-press multi-select for batch deletion
• Attach photos from gallery or camera; persist under app storage
• Share text and optional image via Android ACTION_SEND
• Connectivity banner when offline — the app remains fully usable

Stack

Flutter (Material 3) · sqflite for local persistence · image_picker, share_plus, connectivity_plus, flutter_slidable, flutter_animate

Overview

Drive (ZNL-DRIVE) is a SvelteKit file workspace with Postgres metadata and bytes on Local filesystem or Tigris object storage. It covers the product surface of a small cloud drive: Home library, Shared, Trash, team drives, public links, chunked uploads, and a REST API usable with session cookies or developer API keys.

Capabilities

• Browse folders with breadcrumbs, color labels, pin, and star
• Upload (including chunked large files) and download files or folder ZIPs
• Share with signed-in users; Shared view is read-oriented
• Team drives with a fixed storage provider per team
• Revocable public share tokens for anonymous preview/download
• Trash with restore / permanent delete and scheduled purge
• Better Auth (email/password + optional GitHub/Google OAuth)

Stack

SvelteKit 2 + Svelte 5, Tailwind CSS 4, daisyUI 5 · Postgres (Neon) + Drizzle ORM, Better Auth · Local filesystem and Tigris S3-compatible storage · Docker / Fly.io deployment

Overview

A Unity VR game where the player wakes as a captive in an abandoned building and must gather clues, tools, and power to escape. Built with Team CoreX for a group assignment, using Unity 2022 LTS and room-scale style interaction (grab, flashlight, keypad, fuse/power puzzles).

Role & delivery

• Led software development across core gameplay systems
• Collaborated with UX/UI, project management, and QA teammates
• Interaction design for VR grab, tool use, radio, flashlight, and puzzle props

Stack

Unity 2022.3 LTS · C# · VR interaction and scene scripting

Overview

AdaDDAE extends diffusion-scheduled denoising autoencoders for tabular anomaly detection, evaluated on ADBench (57 datasets). The repository is organized as a named-component framework — each piece (noise schedule, sampling, scoring, ablations) has a clear role so experiments and contribution waterfalls stay reproducible on CPU and GPU profiles.

Research themes

• Leak-safe feature tuning (PCA / scaler pipeline)
• Dataset-adaptive and manifold-aligned noise scheduling
• SNR-stratified timestep selection and contrastive pair sampling
• Variance-uncertainty and diffusion-time posterior scoring views
• Ablation ladders and hardware profiles (CPU through 8–16 GB GPUs)

Stack

Python, PyTorch-oriented training scripts · ADBench datasets and YAML experiment configs · CPU and Vast.ai GPU run profiles

Overview

A Streamlit evaluation app for clinical information retrieval. On a shared random sample of MIMIC-IV discharge notes it compares three retrieval stacks — sparse lexical, dense semantic, and a hybrid medical pipeline — with live search, note detail, and a graded 10-query evaluation panel (Precision@3, Recall@3, MAP, NDCG@3). For research use only; not for clinical decision-making.

Systems compared

• System A — TF-IDF + cosine similarity (lexical baseline)
• System B — dense semantic search with all-MiniLM-L6-v2
• System C — hybrid BM25 + MedCPT dual encoders fused with Reciprocal Rank Fusion (RRF)

Stack

Python + Streamlit · MIMIC-IV / MIMIC-IV-Note (PhysioNet credentialed access) · Hugging Face embedding / MedCPT models · IEEE-format portfolio report (LaTeX)