Agentic Model Customization

• Designed and built Agentic Model Customization, a framework in which a coding agent adapts LLMs to deployment constraints (throughput, memory, context length) by writing its own data, conversion, and training code.
• Built its core components: a sandboxed agent environment, a GPU scheduler, and an evaluation system that serves each candidate in vLLM to check capability and deployment limits.
• Produced a 1.7B-model variant with a 95% smaller KV cache through latent-attention (MLA) conversion and hybridization, keeping accuracy on par with the base model at 1.8× throughput.
• Removed 60% of parameters through compression while keeping 90% of base capability and long-context throughput above the base.

Cardille Lab · Remote

• Designed a benchmark for assessing LLMs' geospatial reasoning capabilities, building a pipeline that executes model-generated code, validates answers, and retries with error correction.
• Implemented a probabilistic time-series model that integrates noisy land-cover classifications into a temporally consistent posterior estimate with quantified uncertainty.
• Designed and optimized an Extended Kalman Filter for spatiotemporal modeling of satellite imagery, using gradient-based parameter optimization to improve generalization across diverse geographies.

• Led tutorials and assessed student work for web programming and database courses.

• Contributed to the development of a cloud-based SaaS platform using Microsoft's technology stack, implementing front-end features in Angular and back-end microservices with .NET Core.
• Led the migration of an active microservice database to MongoDB, optimizing indexing to improve performance.