Saavy Singh

• Supported a graduate Deep Learning course of 60+ MS students through office hours, assignment debugging, and technical guidance on model training, evaluation, and PyTorch workflows. • Helped students reason through CNNs, transformers, optimization, and experiment design, turning course concepts into working implementations. • Collaborated on grading and student support for a fast-paced graduate course, giving feedback on correctness, code quality, and reproducibility.

• Assisted with NYU's GSTEM summer program supporting instruction in foundational CS and math topics for a cohort of pre-college students. • Mentored students through programming exercises and problem sets, providing structured feedback to reinforce computational thinking skills.

• Built and maintained Apache Kafka event-driven pipelines to decouple inventory and pricing updates from the booking flow, improving throughput and fault tolerance under production load. • Improved a high-traffic API latency by 60%+ on hotel booking and pricing services through microservice architecture redesign and targeted caching improvements across partner-facing endpoints. • Developed backend services end-to-end (Java/Spring Boot) supporting hotel partner integrations for 1,500+ partners, from API contracts to data store design and deployment on GCP. • Owned on-call responsibilities for production hotel booking systems, diagnosed and resolved latency spikes, cascading failures, and data consistency issues under live traffic. • Collaborated with cross-functional teams to ship features across the full booking lifecycle, including pricing logic, partner API reliability, and internal tooling improvements.

• Built a 5-stage document extraction pipeline (OpenCV, VGG-19, OCR) to parse scanned healthcare forms into structured key-value and table data; cut manual annotation time by 50%. • Designed a confidence-score routing layer in MySQL that auto-forwarded high-confidence extractions downstream and flagged uncertain fields for review, reducing annotation errors by 32%.

• Built a batch indexing pipeline for a 10M+ image dataset using FAISS with checkpointed progress tracking, reducing reprocessing overhead on failures by 90%. • Engineered a benchmarking harness across FAISS, HNSW, and ANNOY measuring throughput, latency, and top-k recall; drove adoption of FAISS IVF+PQ, improving retrieval latency by 40%.