I'm a data practitioner focused on shipping end-to-end ML/AI and analytics systems — from web scraping and ETL to model training and interactive dashboards. I've built scalable scrapers (10K+ records/day), sharded relational stores for fast dedupe/filtering, productionized ML on AWS SageMaker, and delivered Streamlit/React front ends backed by FastAPI services and BI dashboards.
B.S. Data Science & B.A. Economics (Data Analytics & Policy)
Relevant coursework: Data Structures & Algorithms, Linear Algebra, Statistical Data Science, Regression Analysis, Abstract Math & Proof-writing
Engineered a Selenium/BeautifulSoup scraping pipeline automating job-data extraction (10K+ records/day). Designed a sharded relational database + ingestion scripts for millisecond-level filtering and reliable deduplication. Built a React front end with dynamic search/filter, integrating database queries and Tableau dashboards. Deployed a FastAPI microservice exposing subscription endpoints with scalable, well-structured APIs.
Automated the ML lifecycle with SageMaker Projects: data prep, training, evaluation, and deployment via scripted, reproducible pipelines. Applied workflows on an S3-hosted healthcare dataset; evaluated trade-offs to select features and a production-ready deployment strategy.
Built an automated ETL pipeline (53K+ records) with feature engineering & cleaning. Benchmarked Random Forest vs. Decision Tree (R² 0.92 vs. 0.85) and selected the optimal model for deployment. Delivered an interactive Streamlit app with real-time predictions using sliders and dynamic UI.
Performed EDA on a security/anomaly dataset with Pandas/NumPy/Seaborn. Built simple data pipelines and baseline anomaly-detection models in scikit-learn. Created Tableau/BI dashboards to visualize key trends and insights.
End-to-end pipeline: transcribe audio with Whisper, extract keyframes via OpenCV, encode images with CLIP, generate summaries with GPT-4. Web demo lets users submit a YouTube URL and view the generated summary with captions and keyframes.
Compared open-world detectors, computing IoU/F1 per image; demonstrated YOLO-World's ~2% F1 lead over GroundingDINO (0.73 vs. 0.71).
More exciting projects in the pipeline...
You can reach me at axaanand-at-ucdavis-dot-edu
