A Node/TypeScript backend powering a four-sided marketplace: customers, two kinds of merchants, and delivery riders, all converging on one shared platform for payments, notifications, and dispatch. Each role has its own auth flow and order lifecycle, but every transaction has to stay consistent across all of them.

The interesting engineering wasn't the happy path — it was everything that breaks when money and physical goods are involved. Payment splits between merchants and the platform that have to land in the right wallets every time. Order state machines that need to survive crashed apps, dropped sockets, and payment webhooks arriving out of order or arriving twice. Ledgers that have to reconcile cleanly after a partial refund. Media buckets that slowly fill with orphaned uploads from abandoned listings.

Most of what I'm proud of in this project isn't visible to the end user. It's the recovery scripts that reconstruct orders from payment provider records when something fails midway. The scheduled cleanup crons that keep storage costs honest. The careful state transitions that make sure a rider disconnect mid-delivery doesn't leave the customer hanging or the vendor unpaid. The kind of work that only matters when things go wrong — which, in production, is constantly.

My MSc thesis project at Big Academy UAE / Euclea Business School. The research side trains and compares three CNN models from scratch — one on FER2013, one on CK+, one on AffectNet — to measure how well each generalises across datasets. No transfer learning, no pretrained weights, just standard convolutional architectures to establish honest baselines.

The applied side is a full-stack web app. Upload a photo, point your webcam, or drop a video, and the system detects faces and classifies emotions in real time. The backend runs on FastAPI with OpenCV for face detection and TensorFlow for inference. The frontend is a React app with pages for each input mode, a research overview, and an interactive emotions reference.

The point wasn't to beat state-of-the-art benchmarks — it was to understand where simple models break down across different data conditions, and to build something usable that demonstrates the findings.

Most Discord music bots are glorified search boxes — you give them an exact title, they play it. This one treats requests the way a friend would: you describe a mood, a fragment of lyrics, or a half-remembered song, and a Gemini agent figures out what you actually want.

Under the hood, each Discord server gets its own MusicAgent instance that owns the voice connection, audio player, and queue. Slash commands that are unambiguous (/skip, /pause) bypass the LLM entirely to save tokens — only the fuzzy requests go through Gemini. The agent can clarify ambiguous asks, suggest tracks for a mood, curate full playlists on demand, and politely refuse non-music requests without wasting a token on them.

It's self-hosted on purpose: YouTube blocks datacenter IPs, so the bot runs on a home server or a Raspberry Pi. A deliberate constraint that kept the scope honest.