Darsh Kanjani

BSc dissertation supervised by Dr George Theodorakopoulos. Built a leakage-aware ML pipeline to identify IoT devices from passive network traffic using NFStream-extracted flow features across UNSW-DI, UNSW-AD, and YourThings. Designed a three-tier evaluation framework covering within-dataset, cross-environment, and cross-dataset generalisation. Key result: models reaching 0.72 macro F1 within-dataset collapsed to 0.13 under cross-dataset transfer, showing that standard flow features can behave as fragile environment proxies rather than stable device signatures.

Backend lead for a seven-person team building a financial literacy app. Designed and built the FastAPI backend, including JWT authentication with Argon2 password hashing, Pydantic validation, a deterministic stress-test simulation engine, goal planning, admin-gated endpoints, cascading user data deletion, and a RAG-based advice pipeline using FAISS with sentence-transformers embeddings and an OpenAI LLM for generation. Defined API contracts so the React/TypeScript frontend could develop in parallel.

Solo-built Python platform for personality analysis on Discord, reaching roughly 1,000 users across 100+ servers. Features multiple analysis dimensions, user matching, privacy and consent controls, subscriptions, and an OAuth2 web dashboard. Core analysis uses embeddings, cosine-similarity vector matching, and z-score normalisation. Built with Flask, Celery, discord.py, MySQL connection pooling, and deployed on Linux with systemd and Gunicorn.

Team project building an accessible smart doorbell system. Led backend development and hardware integration using FastAPI, MariaDB, SQLAlchemy, and Uvicorn. Implemented OpenCV facial recognition on a Raspberry Pi, serial communication with an Arduino for servo-controlled unlocking, Firebase Cloud Messaging for push notifications, Ngrok tunnelling for remote access, JWT authentication, role-based permissions, TOTP-based 2FA, and face profile management through the API.

Multi-threaded Python tool for monitoring Discord username availability. Uses Selenium browser emulation, TLS fingerprinting via curl-cffi, proxy rotation, persistent sessions, and Discord webhook alerting. Designed for reliability across extended monitoring periods.

CNN-based classification pipeline for lung, breast, and skin cancers across multiple imaging datasets. Built a Python workflow for image loading, preprocessing, train/test splitting, model training, and reproducible evaluation.