About the job
At NVIDIA, we’re building GeForce G-Assist — an on-device AI assistant that combines Small Language Models (SLMs), retrieval systems, and hybrid cloud capabilities to deliver responsive, context-aware assistance inside the GeForce ecosystem. We work closely across engineering and product teams to ensure G-Assist performs reliably in real-world scenarios.
Responsibilities
Evaluate and improve Small Language Models used in GeForce G-Assist, with an emphasis on accuracy, robustness, and conversational reliability. Identify and mitigate conversation and context contamination, including state drift, prompt leakage, and retrieval cross-talk.
Work with SLM and VLM architectures to support text and multimodal interactions. Collaborate on hybrid architectures that combine local SLMs with cloud-based models.
Optimize local inference using llama.cpp, including quantization, memory usage, and performance tuning. Read, write, and optimize C/C++ code in performance-critical paths.
Design and integrate retrieval-augmented generation (RAG) systems that ground responses in system and user context. Support agentic AI workflows, enabling planning, tool use, and multi-step execution.
Qualifications
Minimum
8+ years of validated experience in system software or a related field, with an M.S. or higher degree in Computer Science, Data Science, Engineering, or a related field (or equivalent experience).
Strong ability to read and write C/C++ code in systems-level or performance-sensitive environments, along with proficiency in Python.
Hands-on experience with llama.cpp or similar local inference frameworks.
Hands-on experience evaluating Small Language Models, including task-based and conversational testing, with an understanding of conversation dynamics, long-context behavior, and contamination challenges.
Knowledge of SLM and VLM architectures and their trade-offs, experience with retrieval technologies and language-model integration, and familiarity with agentic AI patterns such as tool use and planning.
Preferred
Experience contributing to language or multimodal models that power user-facing products, features, or workflows.
A track record of collaborating with product, platform, or systems teams to balance model capability, performance, and user experience.
Demonstrated ability to translate user needs or feedback into measurable improvements in model behavior or system reliability.