Projects

  • In-Memory Computing with Domain Wall Memory

    Jan 2025 – PresentOngoing

    An in-memory computing architecture for domain wall (racetrack) memory that computes directly on stored data, cutting the data movement between memory and processor that dominates energy in machine learning workloads.

    • Uses multi-domain magnetic tunnel junction reads to evaluate several stored operands in a single memory access.
    • Reduces many operands to one result in peripheral CMOS logic, without writing intermediate results back into the racetrack.
    • Supports addition, multiplication, and multiply-accumulate on shared reduction hardware, targeting neural network inference.
    • Evaluated from circuit to system: LTSpice for cell-level energy and latency, NVSim for array-level cost, and RTSim for end-to-end workloads.
    • LTSpice
    • NVSim
    • RTSim
    • Python

  • Vehicle Lane-Change Prediction

    Dec 2023Completed

    My Master's project at Penn State: a hybrid LSTM-HMM model that predicts whether a vehicle will change lanes left, keep its lane, or change lanes right from its recent trajectory, evaluated on the NGSIM highway dataset.

    • Pairs a hidden Markov model with an LSTM that produces its emission probabilities, so the HMM captures maneuver phases while the LSTM learns features from the raw trajectory.
    • Trains one model per maneuver and classifies a trajectory by which model assigns it the highest likelihood, computed with the forward algorithm in log space.
    • Outperformed standalone LSTM and neural-emission HMM baselines trained on the same data.
    • Python
    • PyTorch
    • NumPy
    • Jupyter
    GitHub

  • Skin Infection Detection

    Apr 2023Completed

    A real-time skin infection detection app that classifies photos of affected skin, supporting early diagnosis and reducing reliance on in-person consultations.

    • Built and deployed with a distributed full-stack architecture: an Android client, Flask and Node.js services, and a PostgreSQL database, containerized with Docker.
    • Runs image classification on-device for real-time results.
    • Android
    • Flask
    • Node.js
    • PostgreSQL
    • Docker

  • AI Skincare Advisor

    Apr 2020Completed

    My final year project at HKU: a mobile app that scores a user's skin concerns from a selfie and recommends skincare products suited to them.

    • Built a computer vision pipeline that locates facial landmarks, then scores acne, wrinkles and crow's feet, dark eye circles, sallowness, and skin tone, alongside age and gender estimation.
    • Recommends products from a catalogue compiled from Sephora's website, matching the skin scores and the user's text input with NLP techniques.
    • Served the pipeline through a Flask API consumed by a React Native mobile app.
    • Python
    • OpenCV
    • dlib
    • CNTK
    • NLTK
    • Flask
    • React Native
    GitHub

  • Web-Based Chat Application

    Nov 2017Completed

    A real-time chat system where multiple users can send messages simultaneously, with conversations updating automatically.

    • MongoDB
    • Express.js
    • Angular
    • Node.js