Daivya Shah

• Automated CapEx/OpEx forecasting for goal planning by building an internal OKR platform with Golang microservices and FastAPI on Kubernetes (EKS), using gRPC for inter-service communication and surfacing $120K+ in projected annual labor cost savings
• Built a Real Estate Stacking Plan by standardizing 10K+ leasing records in Databricks using Spark and Delta Lake, enabling 300+ brokers to visualize floor-by-floor occupancy via a Redis-cached React/TypeScript interface, reducing query latency by 30%
• Engineered a data ingestion pipeline with Playwright, Apache Kafka, and LangChain to extract and standardize data from 70K+ property listing URLs through multi-path scraping and LLM-based schema mapping, producing structured JSON for search and analytics

At Ubiq, I worked as a Product Management Intern where I focused on improving user experience and overall product performance. I led user research by conducting 60+ customer interviews and turned those insights into 5 key UX improvements that directly enhanced usability and engagement.

I also executed full-cycle product testing, identifying 50+ bugs and collaborating with engineers to resolve 4 critical issues affecting core functionality. Additionally, I managed an agile backlog of 350+ features, UI enhancements, and bug fixes in Jira, prioritizing development based on user needs and business impact.

• Developed a streaming pipeline to ingest customer call transcripts from Five9 into Snowflake using AWS Kinesis and Glue, enabling EverAI's NLP QA scoring models to automatically audit 78% of a healthcare client's 50K+ monthly calls, up from 12% manual review
• Extended the AWS Glue ingestion pipeline in Python with Protected Health Information (PHI) detection, automatically scanning and redacting 18 HIPAA identifier categories before loading to Snowflake, achieving zero compliance violations
• Built an XGBoost workforce demand forecasting model in Python using features engineered with dbt in Snowflake from Workday HCM and call detail records, reducing staffing error (MAPE) by 11% and cutting WFM planners' manual reporting by 5+ hours weekly

• Fine-tuned a YOLOv8 computer vision model for object detection on 3,000+ labeled crane camera images for 2 underperforming classes via mosaic augmentation and class-weighted loss, improving mAP@0.5 by 0.23 and misclassification rate from 22% to 5%
• Built an anomaly detection pipeline on 2.4M+ crane PLC records across 14 sensor channels using Isolation Forest and LSTM Autoencoder (scikit-learn, PyTorch), detecting 11 failure-risk anomalies from 6 weeks of historical sensor data with 91% precision