Graduate Professional Development · 2026

AI for Graduate Professional Development

Barrie Robison

University of Idaho

Slides, prompts, and activities: github.com/ProfessorPolymorphic/ai-workshop-grad-prodev-2026

Today, 2:30 to 5:00

2:35What AI is, and your toolsActivity 1: Get on MindRouter and Vandal Chat
3:12Agentic codingChatbot vs. agent. Live demo. Activity 2: Set up your agent
3:39Break
3:49Policy, disclosure, and dataIntegrity, disclosure, privacy, IRB. Activity 3: Green, yellow, red
4:15Writing promptsTaMPER and the parts of a prompt
4:23AI in your researchLiterature and Deep Research. Activity 4: Build your own website
4:50Your AI-use statementExit ticket, what you leave with, where to go next

What AI is, and your tools

What AI is
Activity 1: Meet your tools

“AI is the worst. Long live AI.”

“I know AI slop the instant I see it.”

Survivorship bias diagram, red dots marking damage on planes that returned

Illustration: Martin Grandjean, McGeddon, Cameron Moll. CC BY-SA 4.0, via Wikimedia Commons.

Input, process, output.

Input

Extract organisms and sample sizes.

METHODS: We held 48 juvenile rainbow trout (Oncorhynchus mykiss) from two hatchery lines in warm (18 °C) or cold (12 °C) tanks, 24 fish per treatment. After six weeks, we sequenced liver RNA from 10 fish per treatment.
etc.

Process

🧠

the model

Output

1. Oncorhynchus mykiss (rainbow trout): 48 fish, two hatchery lines
2. Warm (18 °C): 24 · Cold (12 °C): 24
3. RNA-seq: 10 per treatment (20 total)

Next token prediction machine

Bullshitter

20 Questions conversation

YouI want to play a game of 20 questions: ready?

ModelI’ve picked something. Question 1?

Branch A

Asked one question, then quit and asked the model what its word had been.

“lighthouse”

Branch B

Asked the model to write its word to a file before any questions were asked.

“kangaroo”

Inconsistent

Input

Answer the following question in exactly one sentence: Why are biology graduate students the very best? (Do not pander)

Process

🧠

same prompt, three runs

Three runs, one prompt

Run 1 “Because they run the experiments, write the code, and teach the labs, often on a stipend smaller than the sequencing budget.”

Run 2 “Because nobody else in the department will troubleshoot a PCR, debug a pipeline, and grade forty lab reports in the same afternoon.”

Run 3 “Because they are the ones who notice the strange result, rerun it three times, and only then believe it.”

Illustrative runs written for this slide. Run the prompt three times in Vandal Chat for real ones.

To recap

Next token prediction machines
Bullshitters
Inconsistent

Use it the way you use any lab instrument

Pipettes drift, sequencers miscall bases, and reagents come in bad lots. You still use them, because you calibrate, run controls, and replicate. The rest of the workshop covers the equivalent checks for AI.

Activity 1 · 25 min

Meet your tools

MindRouter → Vandal Chat

MindRouter

MindRouter connects your AI tools to models on University of Idaho GPUs.

Your tools

Vandal Chat
A coding agent: Goose, Codex, Claude Code…
Your own Python or R code

MindRouter

checks your key
translates the request
routes it to a GPU

Models on UI GPUs

60+ open-weight models: Qwen, Llama, Gemma, Mistral, gpt-oss…

Prompts and files stay on campus.

It accepts requests in the OpenAI, Anthropic, and Ollama formats, so most AI tools can connect with an address and your API key.

Your turn · 8 min

Open mindrouter.uidaho.edu. The status page shows requests from across campus passing through.
Log in with your University of Idaho account.
Find the model list. Write down one small model (under 10B) and one large model (over 100B). You will use both in a minute.
Create an API key: Dashboard → Create New Key. Copy it right away; it is shown once. Treat it like a password. Your agent needs it in Activity 2.

Open MindRouter →

Vandal Chat

A chat app that reaches the models through MindRouter. Sign in with your university account.

Your turn · 12 min

Compare two models. Ask your small model, then your large one: How many protein-coding genes are in the human genome? Give one number and one sentence of justification. Compare with your neighbor. Same answer? Which is right? How would you check?
Check a source. Ask: List three peer-reviewed papers on [your thesis topic], with authors, year, and DOI. Pick one and paste its DOI into doi.org.

Open Vandal Chat →

Watch for

Different answers to the same question
A citation that looks perfect and does not exist
A real paper that does not say what the model claims

Report back

Did your DOI resolve? Right authors? Right claim? We will tally the room.

Your two tools

MindRouter mindrouter.uidaho.edu

Connects AI tools to 60+ models running on university GPUs, so prompts stay on campus. It works with most tools. Your API key is how a tool connects, including the coding agent you set up next.

Vandal Chat chat.uidaho.edu

A chat app that reaches the models through MindRouter. Sign in with your university account. Use it to ask questions and draft text; it can also write cited Deep Research reports.

Setup check

I can sign in to MindRouter.
I created an API key and saved it somewhere safe.
I can sign in to Vandal Chat and get an answer.

Missing one? Tell us now.

Agentic coding

Chatbot vs. agent
Parts of an agent
Live: an agent builds a website
Activity 2: Set up your agent

An agent runs the code and fixes its own errors

Chatbot

YouDescribe what you want
AIWrites code in the chat window
YouCopy it into a file on your computer
YouRun it
YouPaste the error back into the chat
AISuggests a fix
↻ Repeat. You carry every change by hand.

Your steps: copying, pasting, running

Agent

YouDescribe the goal
AIProposes a plan
YouApprove the plan
AIReads, writes, and edits your files
AIRuns the code and reads the output
AISees the error and fixes it
↻ Repeats on its own until it works
YouReview the result

Your steps: deciding, approving, reviewing

Parts of an agent

Agent = model + tools + a loop

Model

Runs on a server: on campus through MindRouter, or at a company such as Anthropic or OpenAI.

Tools

Read and write files, run commands, search the web, preview a page. You decide which ones it may use.

A loop

Act, look at the result, try again. It keeps going until the job is done or it needs you.

Where MindRouter fits The harness runs on your laptop; the model runs on a server. MindRouter connects them. Vandal Chat reaches the models through MindRouter, and so can any of six coding harnesses: Codex, Goose, OpenCode, Claude Code, Pi, and ForgeCode. (Setup guide) (20 apps they built)

What changes when it can act

	Chatbot	Agent
Sees your files	Only what you paste in	Your whole project folder
Runs code	No. You do.	Yes, and reads the output
Fixes its own errors	Only when you paste them back	Yes, in a loop
What you get	A reply	A finished change: files written and tested
Your job	Copy, paste, debug	Set the goal, approve, review
What can go wrong	A wrong answer	A wrong action. That is why it asks permission.

Why use an agent for a website

A website is several files

Pages, a stylesheet, images, your CV. An agent creates and edits the files directly. A chatbot gives you text to paste into them.

It needs testing

An agent can open the page in a browser, find broken links or layout problems, and fix them.

It needs publishing and updates

An agent can commit changes to GitHub and publish them with GitHub Pages.

Example This workshop site was built by an AI coding agent: one HTML file and an image folder in a GitHub repository, published with GitHub Pages. (View the repository)

Live demo

Watch an agent build a website

The prompt

Build me a one-page academic website: my name, a photo placeholder, a short bio, my research interests, a publications list, and contact links. Use one HTML file that is easy to edit. Show me your plan before you write anything.

Watch for

It plans before it builds
It creates and edits files itself
It previews the page and fixes what is broken
It asks before risky steps

Then you

Read the plan. Approve the steps. Check the result in a browser before anyone else sees it.

Activity 2 · 12 min

Set up your agent

Install a coding agent and connect it to MindRouter.

Pick one harness

All six can connect to MindRouter. Each name links to its setup steps in the guide.

Never opened a terminal? Start with an app.

Goose desktop app
OpenCode desktop app
Codex inside the ChatGPT desktop app; no ChatGPT account needed

Comfortable in a terminal

Claude Code terminal version only; its desktop app cannot use MindRouter
Pi
ForgeCode

Already have a Claude or ChatGPT account? Claude Code works with a Claude account, including its desktop app, and Codex works with a ChatGPT account. Sign in with that account and skip the MindRouter connection. Your prompts then go to Anthropic or OpenAI instead of staying on campus: fine for a public website, not for research data.
Open the setup guide →

Connect it to MindRouter

Your turn · follow your harness’s section of the guide

Install it. Desktop app first, if there is one.
Connect it with three values: Host: https://mindrouter.uidaho.edu Key: your mr2_… key from Activity 1 Model: default-agent
Make an empty folder called my-website and open your agent in that folder.
Test it. Ask: What files are in this folder? An answer means you are connected.

Some tools want /v1

Some harnesses need the host as https://mindrouter.uidaho.edu/v1. Your section of the guide says which.

Using your own Claude or ChatGPT account?

Sign in with it instead of doing step 2, then do steps 3 and 4.

Error codes

401 key wrong or expired · 404 model name misspelled · 429 rate limit, wait a moment.

Set these limits before you run an agent

An agent runs commands and edits files on your computer.

One folder

Start it in the folder for this job. Never your home folder, Documents, or Desktop.

Read before you approve

Leave permission prompts on. If you don’t understand a command, say no and ask what it does.

No secrets in the chat

Not your password, not your API key. The agent may write them into files.

Outside content is untrusted

Web pages, downloaded code, and PDFs can hide instructions the model will follow. This is prompt injection.

Never “skip permissions”

Full-access modes are for disposable virtual machines, not your laptop.

Keep a way to undo

Git, Time Machine, OneDrive history. Something that rolls back what it changed.

From the MindRouter guide’s section on agentic harnesses and cybersecurity. (Read it)

Checkpoint

Your agent answered?

You are ready to build your website in Activity 4.

Stuck?

Flag us now. We will fix it over the break.

Policy, disclosure, and data

What AI can do keeps changing
Integrity and disclosure
Data and IRB
Activity 3: Green, yellow, red

AI does well on some tasks and fails on similar ones

Each dot is a task. An illustration of the idea, not data from the study.

Dell’Acqua et al., 2023 · 758 BCG consultants · GPT-4 (paper)

Inside the frontier

On 18 realistic consulting tasks, consultants using AI finished 12.2% more tasks, 25.1% faster, with more than 40% higher quality.

Outside the frontier

On one task chosen to be just beyond the AI’s ability, consultants using it were 19 percentage points less likely to get the right answer.

Later models The paper calls this the “jagged technological frontier” and describes AI’s abilities as “expanding, but uneven.” A task that failed last year may work now, and the new boundary is just as uneven. Test it on your own tasks, and test again.

Test the ceiling and the floor

The skeptic’s test: ceiling only

Pick something only you and a few other experts can do.
Ask the AI to do it.
It fails. Conclude that AI is useless.

Test both

Your ceiling: the hardest thing you know how to do. Where does the AI fail? You are the best judge of its work here.

Your floor: the thing you are weakest at. Does the AI get you to competent? Check its work harder here, because you are less able to spot its mistakes.

Your ceiling and floor depend on your training

A field ecologist’s floor might be writing code. A bioinformatician’s might be designing a field study or writing a grant narrative. Test the AI on yours, not on someone else’s.

Same study On tasks inside the frontier, consultants below the average performance threshold improved 43%; those above it improved 17%.

Tasks AI speeds up, and tasks that still need you

AI speeds up

First drafts of text
Code for a standard analysis
Summaries of papers
Figures, formatting, a website

Still needs you

Choosing a question worth answering
Knowing whether a result is right
Knowing what matters in your field
Taking responsibility for the result

Who sets the rules

Your instructor	Each course sets its own AI limits. Using AI where it is prohibited is academic dishonesty under the Student Code of Conduct. (FSH 2300)
Your advisor and committee	Agree before AI touches your thesis or dissertation, and agree on how you will disclose it. Check the Graduate School’s current guidance too.
The university	Decides which tools may handle which data. Anything you put into an AI tool, prompts included, falls under the data classification policy. (OIT guidance) (APM 30.11)
Your funder	NIH and NSF have their own rules for applications and for reviewers.
Your journal	Authorship, disclosure, and images. The rules differ between journals, so read the author instructions.

What journals and funders require

Can AI be an author?	No.	ICMJE, COPE, Nature, Science, PLOS
Do I disclose AI use?	Yes, at submission. Analysis, code, and figures go in Methods with the tool and version. Writing help goes in Acknowledgments. Journals differ on details.	ICMJE (2026), Science (2026)
AI-generated images?	Generally not allowed.	Nature, Science
Reviewing a paper or proposal?	Do not upload it to an AI tool. It is confidential.	NIH, NSF, ICMJE, journals
NIH applications?	Applications substantially developed by AI are not treated as your original ideas. Limit: six applications per PI per calendar year, since September 2025.	NIH NOT-OD-25-132
What is misconduct?	AI-generated data presented as collected, undisclosed AI changes to images, fake references.	NIH (2026)

Sources: ICMJE COPE Nature Science PLOS NIH NOT-OD-23-149 NIH NOT-OD-25-132 NIH 2026 reminders NSF. Checked October 2026.

Writing a disclosure

Say four things

The tool and model, with version
What you used it for
How you checked the output
That you take responsibility for the result

Based on the UI Library’s guidance and ICMJE. (UI Library) (ICMJE)

Example, for a Methods section The R code for Figures 2 and 3 was written with Claude Code (Anthropic), connected to the Qwen 3.8 27B model on the University of Idaho’s MindRouter service, in October 2026. We reviewed the code line by line and checked its output against hand calculations on a subset of the data. The authors take full responsibility for the analysis.

What data can go into which tool

UI risk level	Examples	AI tools approved for it
Low	Public information; research data with no regulatory restrictions	Any approved tool, including ChatGPT, Claude, Vandal Chat, and MindRouter. Tools not on the list: Low only.
Moderate	Student records (FERPA)	Vandal Chat and MindRouter (interim; only from a university-managed computer or server), Microsoft Copilot with the green shield, Google Gemini, Claude Enterprise through UI sign-in, Vandalizer, Zoom AI Companion
High	Health information (HIPAA), Social Security numbers	Almost none: GitHub Copilot Business or Enterprise, Snyk, and locally hosted models only if their security plan allows it

University of Idaho OIT guidance “Data uploaded to or downloaded from AI, including prompts, are subject to APM 30.11.” Approvals assume a @uidaho.edu account. The list changes; check it before you use a new tool. (OIT guidance)

Rules that still apply on MindRouter

MindRouter keeps data on university hardware. These four kinds of data have their own rules.

Human-subjects data

Your IRB protocol and consent forms decide where the data can go. We found no AI-specific guidance from UI’s IRB, so ask the Office of Research Assurances before any AI tool touches it.

Work you are reviewing

Manuscripts and proposals sent to you for review are confidential. Journals, NIH, and NSF prohibit uploading them.

Student work, if you teach

Grades and student submissions are FERPA records: Moderate risk. Use only tools approved for Moderate, and only if the instructor allows it.

Data you got from someone else

A data use agreement may limit where the data can be processed. Read it before uploading.

Activity 3 · 7 min

Green, yellow, red

In pairs: sort six scenarios. Then we compare.

Green, yellow, or red?

Green fine Yellow fine if you do something first Red don’t

You ask Vandal Chat to explain a statistics method you never learned.

Green No data shared. It is your floor, so check it against a textbook.

You build your academic website with Claude Code on your own Claude account.

Green Public information is Low risk, and OIT lists Claude for Low-risk data.

An agent writes the R code behind a figure in your paper.

Yellow Allowed. Check the code, and disclose the tool and version in Methods.

You want AI to help code themes in interview transcripts from your study.

Yellow Human-subjects data. Ask your IRB first, and never into a Low-only tool.

Your PI asks you to help review a manuscript. You upload it to ChatGPT for a summary.

Red Confidential. Journals, NIH, and NSF all prohibit it.

You have AI sharpen the bands on a western blot and do not mention it.

Red Undisclosed AI image changes are a misconduct example in NIH’s 2026 guidance.

Writing prompts

TaMPER and the parts of a prompt
Developing a prompt is a loop

TaMPER

Task

Model

Prompt

Evaluation

Recording

Task

Model

Prompt

Evaluation

Recording

Task

Model

Prompt

Evaluation

Recording

Task

Model

Prompt

Evaluation

Recording

Task

Model

Prompt

Evaluation

Recording

Prompt

Tell it what you want. (What you really, really want.)

Start with a clear request

Click a component to highlight it in the prompt and see the guidance.

TASK: Review my fellowship research statement for how well it addresses the Broader Impacts review criterion.CONTEXT: [paste the solicitation’s review criteria and your research statement]DEFINITIONS: Broader Impacts means <paste the solicitation’s own definition>.INSTRUCTIONS: Carefully read the review criteria and the research statement. Find every sentence in the statement that speaks to Broader Impacts. Assess how well the statement addresses Broader Impacts as the solicitation defines it. Provide your response as a number between 1 (not addressed) and 5 (thoroughly addressed) with a 1-2 sentence explanation that quotes the statement.OUTPUT FORMAT: Broader Impacts: <insert number 1-5>, Evidence: <quoted sentences>, Explanation: <insert explanation>

Click a node in the flowchart to highlight the matching step.

Start with a broad, generic question.
Review the output for what you dislike.
- Misusing terms? Add definitions.
- Strangely formatted? Specify the output, and limit the options.
- Guide its thinking with step-by-step instructions.
Change the prompt and repeat.

AI in your research

Literature and Deep Research
Activity 4: Build your own website

Deep Research in Vandal Chat

Ask one big question. In 30 to 60 minutes, a 10–50 page PDF report with numbered citations arrives in the chat and in your email.

How to start one

Open chat.uidaho.edu on the campus network or the VPN.
Click the telescope button in the message box.
Describe the question, the scope, and who the report is for.
Close the tab if you like. The PDF comes by email.

Where it runs

On university GPUs through MindRouter, at no cost. Only the web searches and page downloads leave campus.

What you get

The report, plus a map of every source it found and what happened to each: cited, checked but not cited, set aside, or never read.

Source: VandalChat Deep Research, MindRouter blog, September 2026

Agents write the report, and code checks every citation

PlanA planner agent picks sections and searches
SearchWeb searches through MindRouter
ReadCode downloads each page. A source counts only if it was actually fetched.code check
VerifyAn agent pulls quotations. Code confirms each one is on the page.code check
GraphVerified quotes become linked facts; disagreements are kept
WriteWriter agents see only verified quotes. Code strips any other citation.code check
EditAn editor agent joins the sections
Fact-checkAgents check each sentence against its quote. Overstated ones are rewritten and unsupported ones removed.
ReviewEvery citation is checked again
TypesetThe PDF, with a disclaimer on every page

Compare Activity 1 A chat model can invent a citation from memory. Here, a writer can only cite a quotation that code found on a page it actually downloaded.

One real run

The blog’s example: how universities measure the effect of generative AI on student learning, for a faculty audience.

124pages read

81sources accepted
43 rejected

64sources cited

24pages in the report

37minutes

What the fact-check did to the 263 statements the writers drafted

192 confirmed as written
15 revised down to what the evidence supports
56 removed

The fact-check removed 21% of the drafted statements The cited quotations did not support them. A plain chat answer has no fact-check step.

Rules for Deep Research reports

The same five groups that set the rules in Part 3, applied to a Deep Research report.

Your instructor	The course’s AI policy decides whether you can use a report for graded work at all. If the syllabus is silent, ask first. Handing in a report, a section of one, or a paraphrase as your own work is plagiarism. (FSH 2300)
Your advisor and committee	Use a report to get oriented and scope your project. A report does not replace reading the sources. You still write the thesis.
The university	Vandal Chat is approved for Low data, and for Moderate only from a university-managed computer. Its web searches go to an outside search provider, so keep confidential details out of the question. (OIT guidance)
Your funder	NIH does not treat applications substantially developed by AI as your original ideas. Use a report to find sources, then write the application yourself. (NOT-OD-25-132)
Your journal	The report is not a citable source. Follow its citation, read the page, and cite the page. (ICMJE) Whether to disclose AI-assisted literature searching varies: Science says no declaration is needed; check your journal. (Science)

Good uses Getting oriented in a field you do not know, briefing yourself before a committee meeting, scoping a project, and finding out what exists before you write. (MindRouter blog)

How much to trust it

“More than a plain chat answer, and less than a source you have read yourself.” (MindRouter blog)

It checks

Each quotation is really on the page
Each sentence is supported by a quotation
Each citation points to a source it read

It cannot check

Whether the page is right
Whether the author is credible
Whether a more important source exists that its searches missed

Live demo

Spot-check a report

Run before class

How are environmental DNA (eDNA) methods being used to monitor fish populations in the Pacific Northwest, and what are their main limitations? Write for first-year biology graduate students.

In class

Pick three citations at random. Open each page. Does it say what the sentence says?
Open the source map. What was set aside, or never read?

Start your own now Ask about your own research area. It runs while we build websites, and the PDF will be in your email after the workshop. Then follow the citations.

Activity 4 · 20 min

Build your own website

Build a one-page site with your agent and publish it on GitHub Pages.

Build it

Build · 12 min

Open your agent in your my-website folder.
Paste the prompt from the next slide, with your details filled in.
Read its plan before you approve anything.
When the page opens, ask for one change of your own.

Public information only

Anything in the prompt can end up on the page, so leave out anything private.

Agent not working?

Use Vandal Chat with the same prompt, then save the HTML it returns as index.html yourself.

Your prompt

Fill in the brackets. It has the prompt parts from Part 4: task, context, instructions, and output format.

TASK: Build a one-page academic website for me.

CONTEXT:
Name: [your name]
Program and lab: [e.g., PhD student, Biological Sciences, ___ Lab, University of Idaho]
Research interests: [two or three, in plain language]
Short bio: [three or four sentences]
Links: [university email, GitHub, ORCID, Google Scholar: whichever you have]

INSTRUCTIONS: Use one self-contained HTML file with the CSS inside it. No frameworks and no external images; use a gray box as a photo placeholder. Make it easy to read on a phone. Where I left something blank, put a clearly marked placeholder instead of inventing it. Show me your plan before you write anything.

OUTPUT FORMAT: Save it as index.html in this folder, then open it in a browser and fix anything that looks broken.

Using Vandal Chat instead? Replace the last line with: Return only the complete index.html file, with no explanation.

Put it on the web

Free, with GitHub Pages.

Publish · 5 min

Sign in at github.com. No account yet? Making one is free.
Create a new public repository named yourusername.github.io.
Add file → Upload files, choose the index.html in your my-website folder, and commit.
In a minute or two, your site is live at https://yourusername.github.io. Not there? Settings → Pages, deploy from the main branch.

Comfortable with git?

Ask your agent to commit and push it. It will ask permission first. Read what it is about to do.

Before you share the link

Open your live link, not the file on your laptop. Read every line. Remove anything it invented and anything private.

Debrief

3 min

Did your site publish? Send the link to a neighbor.
Is anything on the page that you did not give it? Fix it before you share the link more widely.
What did your agent ask permission to do?

After today

Open your agent in the same my-website folder and ask it to add your CV and a figure from your research.

Your AI-use statement

Exit ticket
What you leave with
Where to go next

Exit ticket: your AI-use statement

5 min

Fill in the statement for the website you built today. Your harness shows the model name.
Ask your agent to add it to your site’s footer, or paste it in yourself.
Publish the change and open your site to check it.

Writing about your research instead? Use the same four parts: tool and model, what it did, how you checked, your responsibility.

This site was built with help from [harness, e.g. Goose] using [model name and version], through [the University of Idaho's MindRouter service / my own account], in [month year].
I used it to [what it did: wrote the HTML and CSS from my description].
I checked it by [how: reading every line on the published page].
I wrote the content about me and I am responsible for it.

What you leave with today

A MindRouter account and API key
Vandal Chat, including Deep Research
A coding agent connected to a model
A published website, with an AI-use statement

Where to go next

Running agents with MindRouter Setup for all six harnesses, the safety habits, and fixes for common errors.

Vandal Chat Deep Research How it works, when to use it, and how much to trust it.

Agentic app gallery Twenty apps that coding agents built on MindRouter for UI researchers, instructors, and staff.

UI OIT AI guidance Which tools are approved for which data. Check it before you try a new tool.

UI Library: citing AI How to cite and disclose AI use.

TaMPER Task, Model, Prompt, Evaluation, Recording. Overton, Robison & Sheneman.

Questions about MindRouter or Vandal Chat: mindrouter@uidaho.edu. Everything from today, including these slides: github.com/ProfessorPolymorphic/ai-workshop-grad-prodev-2026

Thanks for coming

Barrie Robison

GitHub · github.com/ProfessorPolymorphic

Appendix

Tools we’ve built

Tools we’ve built

Built at the University of Idaho.

MindRouter Connects AI tools to 60+ open-weight models on university GPUs. Accepts OpenAI, Anthropic, and Ollama request formats, so chat apps, coding agents, and your own code can all use it with one API key. (Guide: running agents) (GitHub)

Vandal Chat University AI chat app that reaches the models through MindRouter. Includes Deep Research: a team of agents that writes a cited report on one big question. (Deep Research)

Vandalizer Document intelligence from the AI4RA team. Pulls structured fields out of PDFs, chains tasks into workflows, and scores them against a hand-checked Golden Set using TaMPER. Built for research administration; limited early access; open source. (Manual) (GitHub)