HELIOS Lab, Stanford Medicine
Research Governance in the Age of AI
30+ ethnographic interviews with Indigenous scholars and Tribal community members shaped a set of practical tools for strengthening community oversight of biomedical research as AI changes how data is collected, interpreted, and used. I served as the study’s lead designer, leading much of the interviewing and qualitative analysis and owning design operations and tool development.
At a glance
- Scale
- 30+ ethnographic interviews
- Role
- Lead designer · Researcher
- Methods
- Community-engaged research · qualitative analysis · research governance · digital product prototyping
- Outcome
- Consent-review tool built and presented to the Tribal Health Committee; screening and research-dashboard tools in development; publication subject to Tribal approval
AI is moving fast into biomedical research, but Indigenous Peoples, whose data, land, and communities are directly affected, are often not consulted in how these tools are designed and governed. That problem is what I joined the HELIOS Lab (Stanford Medicine) to work on in my first quarter at Stanford, through AI-READI, a consortium under the NIH Bridge2AI initiative. Our study focused on the enabling conditions for good research practices and the ways Indigenous governance can be strengthened.
Owning the research process
I began by synthesizing literature on AI governance, Indigenous data sovereignty, and community-engaged research, then co-authored our interview protocol with the PhD candidate I work with on this study. Across 30+ interviews with Indigenous scholars and Tribal community members, we studied how AI and machine learning were entering biomedical research, where existing research relationships broke down, and what stronger governance could look like. The study unfolded over two years because trust-building and Tribal governance set the pace.
I built much of the study’s qualitative-analysis infrastructure—codebook, coding rationale, analytic memos, and synthesis—then translated the findings into tools intended to be useful in real research decisions rather than archived at the end of the study. In April 2026, we presented and validated our initial findings with the Tribal Council.
Consent review
Plain-language review of research consent forms, run entirely on Tribally controlled hardware.
Built · presented August 2026
Research insight dashboard
In development · for Tribal Council approval
Partnership screener
In development · for Tribal Council approval
Design principles
These tools are built using a set of five design principles.
Increase common knowledge.
Make the terms of research legible to everyone at the table, not only to the researchers.
Titrate friction.
Add friction where a decision deserves care, and remove it everywhere else.
Imagine what could go right.
Guard against the harms of past research, and leave room for what a good partnership could produce.
Consider relationships as scaffolded.
Trust is built in stages, so tools should support each stage of a research relationship, not a one-time sign-off.
Don’t stop at another PDF.
Translate research into tools people can use while making consequential decisions.
Governance became the technical architecture
One design constraint changed the technology stack completely: sensitive consent materials could not leave Tribally controlled hardware. A hosted model API was therefore the wrong architecture, regardless of convenience. I built the tool on a locally served open-weights model and designed the interface around the reality of slower on-device inference.
I tested a smaller model first. It couldn’t reliably hold the tool’s output format, so I moved to a larger one and accepted the slower speed, then evaluated the tool against a 30-document corpus of real consent forms. To make the wait workable, the analysis streams in section by section: on a real consent form, the first section appears in under a minute instead of after a three-minute wait for the whole thing. Uploaded documents are read in memory and never written to disk.
The point was not to add AI to the process. It was to make the technical system obey the governance principle.
OLMo 3.1 32B Instruct (q4_K_M) · served locally with Ollama · FastAPI · streamed, section-by-section output · no files written to disk
One review, step by step. Everything below the line of visibility happens on Tribally controlled hardware.
- 01 Bring the form
- ReaderUploads a consent form or pastes its text, and can add what they’re worried about.
- Tribal hardwareReads the file in memory; nothing is saved.
- 02 Check
- Readable text?Yes: The review begins.No: a scanned or locked file gets a plain-language explanation, and the reader tries again.
- 03 Read
- Tribal hardwareA local model reads the whole form.
- On screenLays out the 12 questions ahead, with progress.
- 04 Stream
- Tribal hardwareStreams each section as it’s finished.
- On screenFills in section by section.
- ReaderStarts reading within a minute.
- 05 Worth asking
- Tribal hardwareIf needed, a second pass fills in missing questions.
- On screenFlags what’s worth asking, with a question for each.
- 06 Worries
- Wrote a worry?Yes: It’s answered using only the document.No: straight to the reminder.
- 07 Take it along
- On screenA reminder that this is not legal advice, and who to talk to.
- ReaderPrints it or saves a Word copy to bring along.
Where it stands
In August 2026, we shared the consent-review tool with the Tribal Council’s Health Committee. That conversation is shaping the next stage: further testing within the Tribe, on the community’s terms and timeline. A screener and a research dashboard are in development for Tribal Council review, and the white paper will follow additional interviewing later this year.
All outcomes of this research (a white paper, a consent-review tool, a research dashboard, and a partnership screener) are built squarely for the Tribe we’re working with. They have the final say over whether our work is published and shared with the world, corresponding with our core value of data sovereignty and strengthened governance.
Philanthropy showed me the value of place-based work. The d.school and HELIOS gave me the skills to build for the communities I care about.
Due to the sensitive nature of this work, documentation is limited. I’m happy to discuss this project in more detail. Get in touch.