Projects
Empathetic Multimodal Conversational AI
Full-stack web app integrating PyTorch ML models with OpenAI/ElevenLabs APIs for a multimodal AI therapist. Orchestrated multimodal nonverbal cue recognition pipeline for facial and voice affect adapting to emotional status
Recreation of 3D Neural Radiance Field Paper
This project recreated the Neural Radiance Fields (NeRF) paper in PyTorch and NumPy to learn a 3D scene from multi-view images and render novel viewpoints. I implemented the full training pipeline, including custom ray sampling and positional encoding techniques that improved training stability and reduced view-dependent artifacts.
Deep Learning Clothing Semantic Segmentation
A vision based semantic segmentation with ResNet-18 and U-Net on 1k+ clothing images using PyTorch. Trained multi-class and binary models with Adam, monitoring loss and per-class accuracy on validation data. Leveraged CUDA GPUs in Colab to accelerate training and evaluation, achieving 0.48 and 0.045 validation loss.
Emotion Recognition Deep Convolutional Neural Network
Noticed there is no open source emotion detection asset for Unity developers. So created a ML driven real-time facial emotion recognition to Webcam pipeline using the FER+ DCNN and Unity’s Barracuda inference engine. Normalized frames with 68-point landmark alignment to cut pose variance and boost classification stability. Optimized frame processing and model I/O for low-latency inference suited to interactive AR/VR experiences.
Personally Identifiable Information Masking
Token-level NLP system with FFNN and RNNs to automatically detect and redact PII across 10 entity classes. 7-component training stack from preprocessing to evaluation, reducing runtime by 23% and scaling batch inference. Achieved 0.89 weighted F1 on val and improved performance +6 points over baseline for token sequence tagging.
Interactive 3D Weather Visualizer with ChatGPT for AR Wearables
A Spectacles AR application combining geolocated weather APIs with ChatGPT for on-tap weather summaries. Mapped live temperature and conditions to 3D models, icons, and VFX that update in real time with forecasts. Implemented async, gesture-driven pipeline with API caching, error handling, and UI tuning for fast AR on device.
Revenue-Optimizing Auto-Loan Pricing Engine with Random-Forest Uplift Modeling
Built an end-to-end pricing-optimization pipeline for an online auto-lender. Cleaned 40 k+ loan records, explored acceptance-vs-APR trade-offs, and used a logistic baseline to quantify current revenue. Iteratively optimized each customer’s rate, demonstrating a portfolio-wide uplift of ≈ $84 k in expected revenue for the test cohort.