Ashutosh Naik

Nomura Holdings America, Inc. · New York, NY

• Migrating Nomura's credit ETL system (RETL) from Prefect 1.x to 2.x as part of a broader event-based processing (EBP) transition; refactoring monolithic sequential flows into parallel, modular task subflows with Dask-backed distributed execution.
• Architected the "Mario" configuration microservice (FastAPI, PostgreSQL) with Smart Insertion and SQLAlchemy connection pooling; integrated into an event-driven architecture via Stargate hub (Solace PubSub+, Mercury Messaging).
• Led infrastructure-wide migration to AWS Graviton (ARM64); diagnosed and resolved x86→ARM64 integer overflow discrepancies in Basel-compliant risk model codebases, restoring 100% data parity.
• Built a RAG-based credit reporting tool (LangChain, LLMs) with a natural language query interface, automating analysis of complex financial reports.
• Engineered a Cloud Lifecycle service automating EC2 shutdown and CNAME management via REST API, reducing infrastructure overhead by 12%.

University of Colorado Boulder — CAIRO Lab · Boulder, CO

• Primary implementer of ShelfAware (IEEE RAL 2025): designed a real-time semantic particle filter fusing depth and distributional semantic likelihoods (JSD over object-class counts) — achieving 96% global localization success vs. 22% MCL / 10% AMCL at 9.6 Hz on consumer hardware.
• Built ShelfHelp's two-stage detection pipeline (AAMAS 2023): YOLOv5 + frozen autoencoder with cosine similarity, enabling scalable product detection without retraining; MDP-based manipulation guidance achieving 90.66% task success (n=15, blindfolded users).
• Designed an autonomous anomaly detection system: trajectory-based spatial analysis fused with GPT-4V for semantic interpretation; introduced Jensen-Shannon Divergence for shelf stock fluctuation quantification.