Hello, I'm Cat.

Designed and deployed scalable data pipelines using PySpark on GCP Dataproc to ingest, preprocess, and store millions of multi-domain documents daily for LLM training

Developed and benchmarked a high-precision RAG framework, leveraging retrieval outputs to contextualize and curate long-context training datasets

Trained and conducted ablation studies on LLMs across data and architecture configurations using distributed GPU clusters on CoreWeave with Kubernetes and Grafana

Improved Patch-based DDPM framework by incorporating conditional diffusion, achieving a 15% boost in SSIM for PET image reconstruction

Developed two meta-learning algorithms for image classification: DG-SharpMAML and AS-MAML, incorporating gradient-matching and sharpness-aware minimization

Demonstrated performance through ablation studies and theoretical analyses, beating state-of-the-art by 2% on accuracy benchmarks

Enhanced Nvidia's NeMo ASR into a speaker diarization pipeline, achieving word error rate below 5% on videos over 1 hour in length

Built a Streamlit app for video transcription with OpenAI Whisper, integrating Longformer and BART for summarization

Proposed robust training pipeline for YOLOv8 to detect pharmaceutical products using 21,500 synthetic images, achieving 98% mAP@50 score

Optimized PaddleOCR for information extraction from 40,000+ multilingual invoices for SAP clients

Automated data labeling and information extraction with Python, reducing processing time per batch by 16%

Standardized 31 datasets and integrated 106K documents via New York Times API to create 3 new datasets

Streamlined dataset mapping with Python and JavaScript on AWS OpenSearch, reducing upload time by 12%

Categorized 256 application essays using Linear SVC, Naive Bayes, and KNN, achieving 88% accuracy