Partners: eScience Institute, National AI Research Resource (NAIRR) Pilot Program, Schmidt Sciences (Virtual Institute for Scientific Software), Microsoft Azure, AWS, Hugging Face; CloudBank

SSEC Engineers: Landung (Don) Setiawan, Anant Mittal, Cordero Core, Niki Burggraff, Anshul Tambay, Carlos Garcia Jurado Suarez, David A. C. Beck, Andrew J. Connolly, Sarah Stone, and Vani Mandava

Background 

Working within the eScience Institute, SSEC’s research software engineers (RSEs) are embedded across many scientific domains at once—astronomy, earth and climate science, agriculture, and health—giving our center an unique perspective on how AI-assisted workflows succeed and fail in research settings. The goal of LLMoxie is to close the widening gap between AI-led software creation in industry and the distinct needs of academic research environments. 

Software Problem

The rapid maturation of AI coding agents has transformed commercial software development, but these tools are poorly calibrated for science. Trained and benchmarked overwhelmingly on commercial code, off-the-shelf agents ignore the conventions of the scientific Python libraries they invoke, mishandle sensitive or embargoed research data, and leave decision trails that are difficult to reproduce. Additionally, scientists face a overwhelming set of practical questions about AI: Which model should I use? Where should I host it? How do I work with it securely, keep track of costs, and know that it is actually working? A recent study from one of our colleagues describes the current environment: researchers overwhelmingly reach for general-purpose chatbots over tools integrated into their programming environments, and only a small fraction of research repositories combine high AI usage with high software-engineering maturity, raising real concerns for reproducibility and open-science integrity. When considering accessibility and lowering barriers to entry, given the wide array of presently available models and harnesses, easing th process of environment setup and how to choose between each becomes essential.

Software Solution

LLMoxie is an institutional AI Gateway that pairs with RSE-Plugins to form an open-source ecosystem which can provide an onboarding service. The platform’s three-tiered architecture separates concerns so each layer can be adopted independently. An inference layer routes requests transparently across cloud (Azure, AWS Bedrock, Google Vertex AI) and on-premise (vLLM/HPC) hosted models. A control plane built on LiteLLM and MLflow provides a single, governed, OpenAI-compatible gateway with authentication, rate limiting, per-user and per-team budgets, PII masking, request logging, and observability. Finally, an application augmentation layer delivers domain-specific workflows to coding agents, encoding accumulated RSE knowledge in a reusable context for AI coding agents. RSE-Plugins is a Plugin–Agent–Skill hierarchy that packages scientific Python development discipline (packaging, testing, documentation, reproducible environments with pixi), domain-specific knowledge, a structured six-phase research-and-implement workflow that emits auditable reasoning trails, and full project-lifecycle management. 

Due to the fact that every interaction flows through a single governed gateway, LLMoxie is also an observatory: its logging, telemetry, and MLflow-based evaluation capture a rich, consented record of how researchers actually use LLMs: which models, prompts, and workflows they reach for, where agents help, and where they fail. The entire Azure stack can be provisioned declaratively as infrastructure-as-code, and both the platform and plugins are available as open-source software. 

Impact

Drawing on twenty months of practice across a portfolio of active projects and more than twenty collaborators spanning disciplines and roles, LLMoxie repositions large language models from generic code generators into domain-aware collaborators that respect community conventions, handle research data responsibly, and produce auditable provenance of reasoning. Centralized, governed access lowers barriers to entry while modular, reusable skills and plugins let teams fold AI assistance into their own workflows and, in several cases, contribute plugins back to the community. The gateway has already supported upwards of 150 users across workshops, individual scientific software development, and AI-assisted research workflows. 

Beyond serving individual projects, a central mission of LLMoxie is to learn from its own use. With participant consent, every request, cost, model choice, and workflow that passes through becomes data for an ongoing study aimed at understanding how researchers actually adopt LLMs for coding and beyond. Usage logs, dashboards, and structured evaluations let the team see which practices genuinely help, where agents introduce errors, and which domains and tasks are best served, closing a feedback loop in which evidence from real scientific work continuously refines the platform, its plugins, and the guidance offered back to the community. This positions LLMoxie not only as infrastructure that scientists can trust, but as a growing evidence base on effective AI-assisted research. Prioritizing institutional data governance, observability, and ease of access, LLMoxie offers a sustainable foundation for high-quality, AI-augmented scientific research.

The work was accepted for presentation at the ACM KDD 2026 SciSoc Workshop and supported a 60-person workshop at Schmidt Sciences’ 2026 Interdisciplinary Science Summit. Additionally, the open-source RSE-Plugins ecosystem, one-click sandboxes, and tutorials continue to make these practices available to the broader research community. 

Related Repositories

LLMoxie platform 

RSE-Plugins ecosystem