My Projects

Showcasing my work and creative solutions

MorningTide Project Image

MorningTide: AI-Driven Therapy Assistant

December 2025 - Present | New York, NY

An intelligent therapy support platform that combines sentiment analysis with session planning automation. Uses low-latency AI models to analyze client journal entries and provide therapists with actionable insights for effective session planning.

Sentiment Engine: Implemented DistillBERT for emotional intensity analysis (valence scores over 20 emotions), retaining 96% of BERT performance while reducing inference latency 5x.
Content Analysis: Built sentence importance model achieving 95% accuracy in identifying high-impact journal entries for therapy-focused analytics.
Session Planning: Developed RAG indexing module analyzing weekly journals with therapy documents, achieving 60% reduction in manual prep time.
Flask React DistillBERT RAG PostgreSQL SQLAlchemy PyTorch
Full demo coming soon
Best Children Story Generator

Fine-Tuned Small Language Model for Children Story Generation

November 2025 | New York, NY

Fine-tuned Qwen3-1.7B small language model (SLM) for child-safe story generation. Targeted working parents needing quick, safe and creative bedtime stories. Achieved top 4 recognition among 50+ competing teams at AWS & AGI House Build Day.

Model Fine-Tuning: Trained Qwen3-1.7B on TinyStories dataset (2.2M+ stories) using LoRA fine-tuning techniques.
AWS Optimization: Leveraged AWS Trainium and Optimum Neuron for 1-day delivery requirement, balancing speed with stability.
Safety & Reliability: Implemented content safety filters ensuring child-friendly output; surpassed LLMs by 20% in generation reliability via human evaluation.
Small Language Model LoRA Fine-Tuning AWS Trainium Optimum Neuron PyTorch
View Code Proof of Recognition
Credit Card Fraud Detection MLOps Solution

Credit Card Fraud Detection MLOps Solution

July 2025 - September 2025 | Hong Kong, HK

Enterprise-grade, end-to-end MLOps solution for real-time credit card fraud detection. Features comprehensive feature engineering, multi-framework support, and production-ready streaming inference at scale.

Real-Time Pipeline: Apache Spark Structured Streaming processing 10,000+ transactions/second with sub-second latency, improving fraud detection rate by 15%.
ML Automation: Automated pipeline with MLflow achieving AUC-PR of 0.92, 30% reduced training time through optimized feature engineering.
Data Simulation: Confluent Kafka to continuously generate labeled synthetic transactions with 6 distinct fraud patterns, enhancing model robustness by 30%.
Apache Spark Confluent Kafka Apache Airflow MLflow PostgreSQL Docker Redis MinIO Boosting Models
View Code
Deep Learning News Recommendation System

Deep Learning News Recommendation System

July 2025 - August 2025 | Hong Kong, HK

NLP-based news recommendation engine using comparative deep learning architectures. Deployed as scalable Streamlit application supporting 500+ concurrent users with sub-second latency.

EDA & Optimization: Conducted exploratory data analysis on user engagement using NumPy, pandas, and matplotlib; improved dataset quality by removing 20% redundant features.
Architecture Comparison: Evaluated CNN, RNN and transformer-based architecture, achieved best AUC of 0.86 (24% better than CNN baseline).
Production Deployment: AWS-hosted Streamlit application with optimized UI, serving 500+ concurrent users with sub-second prediction latency.
PyTorch CNN/RNN Transformers Streamlit AWS pandas GloVe/Word2Vec
View Code

Interested in Working Together?

Let's collaborate and create something amazing!

Start a Project