AI4RA Tip of the Week #20

MindRouter and the Case for Transparent and Sovereign AI Infrastructure

MindRouter is an open-source AI gateway developed by Luke Sheneman, co-PI of the AI4RA project and Director of the University of Idaho’s Research Computing and Data Services. It sits as a middle layer between tools like Vandalizer and the AI models that power them, routing requests across GPU servers and providing a single, reliable access point for running large language models on university-controlled infrastructure. MindRouter addresses important concerns in research administration: accountability, data security, resource governance, and operational transparency. How do you know who is using AI resources, which models they are using, how much capacity they are consuming, and where sensitive data is being processed? MindRouter gives administrators the tools to answer those questions directly. 

Seeing What AI Actually Uses 

MindRouter tracks token consumption by user and request, making it possible to see how AI capacity is distributed and configure quota budgets for individuals or groups. Each AI server also runs a lightweight monitoring agent collecting live hardware telemetry:  utilization, memory usage, temperature, and power draw.  MindRouter does not currently calculate the exact watts attributable to a specific query, but it provides two complementary forms of visibility: computational visibility (hardware load, memory, power draw, and capacity) and usage visibility (which users and applications are consuming tokens, which models are being used, and how demand is distributed). Together, these provide a far clearer operational picture than a typical hosted AI service, which may offer little more than a monthly invoice. 

Keeping Sensitive Data Secure

Because model inference runs on university-controlled infrastructure, prompts and data sent to locally hosted models never need to be transmitted to an external commercial provider.  This is an important distinction for administrators working with sensitive institutional, student, research, or sponsor information. MindRouter also includes data loss prevention (DLP) capabilities that scan prompts and responses for content such as Social Security numbers, credit card numbers, and export-controlled information, using both pattern-based detection and a named-entity recognition model. Depending on institutional policy, detected content can generate alerts, be redacted, or blocked. Local infrastructure reduces external exposure, while DLP controls catch sensitive information regardless of where the model is running. 

Accountability, Model Control, and Open Source 

MindRouter logs each request with the authenticated user, model used, token consumption, request timing, client information, and source network details. Web searches receive their own audit records. Prompts can be retained according to configured policy; model outputs are not retained as part of the normal audit record. This means the system can document who asked what, which model processed it, and how much capacity was consumed. 

Institutions also control which models are available. Rather than relying on a single vendor-selected model, administrators can choose from a wide range of open-weight models suited to different tasks (reasoning, speed, long context, or specialized research workflows) and define content policies enforced at the infrastructure level rather than left to individual applications.  Institutions can also control the stability of the models, where models perform consistently for long periods of time rather than constantly changing as with the Frontier cloud providers. 

MindRouter provides an OpenAI-compatible interface, meaning tools like Vandalizer and other applications can share the same institutional AI layer for authentication, routing, monitoring, DLP, auditing, quotas, and model governance rather than each implementing its own. 

Finally, MindRouter is released under the Apache 2.0 open-source license. Institutions can inspect, audit, modify, and deploy the source code independently. Transparency is not only the ability to see what the system is doing at runtime; it is also the ability to inspect how the system itself works. 

Why This Matters for Research Administrators 

Research administrators operate in a high-accountability environment where processes must be documented, defensible, and reproducible. AI systems that are difficult to audit or unclear about where institutional data is processed are a poor fit for that environment. MindRouter makes visible what conventional AI services tend to hide: who is using the system, which models, how many tokens, how the infrastructure is performing, and what the GPU fleet is drawing. Sensitive information can be detected through DLP. Requests are attributable through audit logs. Model inference stays on university-controlled infrastructure. 

For research administrators, that combination means AI can be approached with the same principles applied to other institutional systems: access controlled, activity documented, sensitive information protected, policies enforced, resource use measured, and infrastructure independently inspected. The goal is not simply to make AI available, but to also make AI use observable, governable, secure, and accountable. 

How Can I Learn More?

MindRouter — https://mindrouter.ai
AI4RA Technical Office Hour — Every second Tuesday at 11 AM PST (https://ai4ra.uidaho.edu/#events) 
AI4RA Office Hours — Every fourth Tuesday at 11 AM PST, with live Q&A and demos (https://ai4ra.uidaho.edu/#events)