Designing an AI-Powered Research Pipeline

These are the presenter’s own notes for each slide, lightly edited for reading. They carry the parts of the workshop that live in the narration rather than on the slide — the reasoning behind each move, the case-study numbers, and the tools named along the way. Slides 45–54 are appendix and glossary material that was not presented live.

1. 01

   Cover

   Opens on the promise: turn your domain expertise into an AI-powered research pipeline in two hours. No coding and no prior AI experience — for most people in the room, this was a first attempt.

2. 02

   <> Framing

   Section marker. The reframe the whole session turns on: you don’t learn to use AI, you learn to interact with it.

3. 03

   Today

   The mission, stated transparently: this is how one researcher is working through this moment, out loud. Not a finished doctrine — confident about the approach, humble about the certainty.

4. 04

   Why we’re here

   The frame that unifies a bimodal room. Everyone present is an established researcher with hard-won ways of bringing rigor, hit by a fast, mid-career disruption. The common ground isn’t AI fluency — it’s the disruption.

5. 05

   What colleagues want

   Who was in the room: UVA colleagues across a wide range of fields, nearly all already using AI in some form, most wanting a repeatable process rather than one-off tricks. In the pre-workshop survey, finding and synthesizing literature ranked as where the research pipeline loses the most time.

6. 06

   In your words

   Closes the who’s-in-the-room section on colleagues’ own voices, trimmed from UVA faculty free-text survey responses and attributed by field. The throughline: help me do more, but don’t replace my voice, ideas, or reasoning.

7. 07

   Who I am

   Background, and an explicit statement of what the instructor is not. A physician and public-health researcher in the School of Medicine who teaches EM residents, with research on the built environment — how the constructed world shapes health, equity, and well-being. Working alongside architects and designers led to a decade directing the Medical Design Program.

8. 08

   Medical Design Program

   Where the interest became a program: ten years teaching first-year medical students to work like designers, with real clients across UVA Health and Charlottesville. The photographs carry it — the point is a way of moving through ambiguity.

9. 09

   The turn

   The pivot of the talk: the years spent learning to approach problems like a designer — and the facilitation skills for holding a group in an ambiguous space — turn out to be exactly how you work well with generative AI.

10. 10

    The premise

    The premise in one line. Stop treating AI like a new Excel — or SAS, or Stata. These systems are probabilistic, not deterministic. Find your own way of interacting: lean on their strengths, and carefully check the asks that need checking.

11. 11

    Four lessons

    Four lessons learned by doing, across ten years of sprints. One: facilitation plus a heavy capture habit — recordings, transcripts, whiteboard photos, multimodal inputs — is how you pull signal out of a group. Two: design techniques keep the group in problem-exploration space instead of racing to solutions. Three: capture, synthesize, iterate — that loop surfaces the more interesting ideas. Four: getting prototypes out fast is a core tenet of design thinking, and artifacts are now nearly free.

12. 12

    <> Health Design Sprint

    Section marker. The design-sprint method behind the approach, and the lens for everything built during the session.

13. 13

    Case · M4 sprint

    Proof the method works: the M4 Health Design Sprint. Eight medical students with no coding background shipped working clinical tools in ten days. The slide walks the artifacts — what they built, how the room worked, what the process looked like.

14. 14

    Secret sauce

    The method, and the mindset shift at its center: put your creative energy into the what — the outcome you want, the metric you’ll have hit, what you’d love to know — and hand the how to the system. It pulls and structures data, including messy sources, picks techniques, and builds things you’d never have thought to build.

15. 15

    <> Pipeline

    Section marker for the spine of the workshop, with the norms said out loud: free-flowing, name your own tools, one live build through the five stops. Not the only way to do this — just how it has been done in practice.

16. 16

    Master Figure

    The canonical pipeline figure. Input: your domain expertise. Output: a pipeline you own. Five stops — context library, direction, literature, data and analysis, communicate. The loop (diverge, converge, repeat) runs at every stop; verification and your judgment run under all of it.

17. 17

    Define · Context

    Level-set before stop 01. Context is everything the model can see at the moment it answers. It is not memory. The practical consequence sets up the whole pipeline: what you get back is mostly a function of what is in the window.

18. 18

    01 Context library

    Stop one. The simplest version is a stable folder on your computer — drafts, articles, everything about the project. With not much more effort, a GitHub repository: the same idea plus history, sharing, and agent access. Either way it stops being storage and becomes the central nervous system of a live collaboration with AI.

19. 19

    01 In practice

    Papers and drafts are the obvious layer, and not the place to stop. Meeting transcripts, project emails, whiteboard photos, datasets, and prior grants and reviews all make a large difference. The models don’t need any of it cleaned up first — you just need a place to put things.

20. 20

    01 Tips & tools

    Tools used to fill the library. Record inputs instead of typing; Wispr applies light AI editing and formatting to voice notes. Otter.ai and Granola transcribe meetings, and official Claude and ChatGPT connectors make those transcripts part of the context library, or a searchable archive. Snipd clips AI summaries of podcast moments into Notion, which connects easily to the major AI harnesses.

21. 21

    01 Brainstorm

    Facilitated brainstorm on techniques people are already trying, kept open so that a named tool or a shared habit benefits the whole room.

22. 22

    02 Direction

    Stop two, where design thinking comes into its own and keeps the work in first principles. The mindset: there’s no reason not to try building the output right now. Paper, dashboard, or website — render it as an HTML file on day one. Being able to see it is a quick test of whether it’s interesting.

23. 23

    02 In practice

    Two live examples. The Med AI Map: could the published literature map medical AI’s maturity? The data-center report: what would a social-license-to-operate valuation look like? Both were rendered as prototypes early, and both drew feedback within days that the direction — if it could be pulled off — was worth doing.

24. 24

    02 Tips & tools

    Two habits. Ask for outputs in HTML or other interactive, dynamic formats — dashboards and sites, not just documents. And flip the interview: have the AI ask you questions about the research idea, adopting personas such as contrarian, security-focused, or skeptical funder.

25. 25

    02 Brainstorm

    Facilitated brainstorm pushing toward a concrete output. The picture in someone’s head is usually further along than they think.

26. 26

    03 Literature

    Stop three, and the new 800-pound gorilla: agentic systems. Hooked up to connectors and APIs — Consensus, Semantic Scholar, PubMed, OpenAlex — they pull the literature directly into the context window. It still needs verification and validation, including agents checking themselves, but the judgment stays yours.

27. 27

    Case · Med AI Map

    The Med AI Map, and the division of labor behind it: AI wrote the pipeline across roughly twenty logged sessions; the researcher supplied the judgment — the definition of maturity, the scope, the gold standard. In week one, an adversarial AI review caught the AI inventing a headline statistic in the project’s own draft; the standing rule since is that every number must be script-emitted from committed data. Five weeks from first search string to a preliminary map in front of UVA’s AI Strategy Council, 100+ commits, and $4.50 in API credits on a personal plan. 4,968 papers labeled in a single session.

28. 28

    Define · Agent

    An agent is a model with tools and permission to take several steps. Good for multi-step retrieval and checking — you still own the judgment.

29. 29

    03 In practice

    The Med AI Map corpus. The question: how mature is medical AI, according to the literature itself? The system was allowed to suggest the approach — read the field at scale. Several hundred thousand abstracts and full-text articles were pulled into the project as a routine step, something that had not seemed possible not long before.

30. 30

    03 Tips & tools

    Most major literature databases connect directly to Claude or ChatGPT — PubMed, Semantic Scholar — by API or other connection types. Consensus is an AI-powered literature review tool, useful for first-pass large-scale synthesis and reference validation. CORE at Oxford allows full-text article search, often accessible with UVA credentials.

31. 31

    03 Brainstorm

    Facilitated brainstorm. This one tends to surface the trust question, and verification practices are worth trading across fields.

32. 32

    04 Data & analysis

    Stop four: going out and getting the data, and setting up the analysis. Scraping datasets and unstructured sources, parsing with natural language processing, agents standing up techniques that would previously have been outsourced — with verification built in. They check themselves; you check them.

33. 33

    04 In practice

    Data-center operations events: scattered, unstructured, never in a tidy dataset. Understood in theory, but previously not technically capturable. Now scraped, parsed with NLP, and in the analysis — a routine step inside the pipeline.

34. 34

    Case · Data centers

    The data-center work, led by a reframe: the question was not unanswerable, it was unassembled — every fact already sat in a public record, and AI collapsed the cost of turning records into data. The 266-project census went from “we should count the denominator” to a shipped, source-hashed dataset in a day. The second lesson: the same agents, run adversarially against their own output, caught inflated model metrics and duplicate events. Speed is what made rigor affordable, not what was traded away for it — a six-lens audit never happens at solo scale otherwise.

35. 35

    04 Tips & tools

    Agentic auto-research loops — runs that plan, execute, and iterate without a human at every step. And adversarial agentic review: a second agent set against the first one’s output.

36. 36

    04 Brainstorm

    Facilitated brainstorm on written-off data. Someone always has a source they assumed was out of bounds.

37. 37

    05 Communicate

    Stop five. Don’t ask it to write the paper — brainstorm the detailed outline together. Set up reviewer personas that hold a view rigidly and push back; the thinking stretches, and the draft gets better much faster. Still your voice, your ideas, your judgment — exactly the line the room drew in the survey.

38. 38

    Define · Artifact

    The shift is asking for the object — the paper, the dashboard, the site — rather than a description of it. Prototypes are nearly free, so the ask should be concrete.

39. 39

    05 In practice

    The same partner works the funding side. As an idea forms, have it look for funding sources and where the work might fit. Use it to define the outcome from the start, honing the thinking toward specific goals, and keep part of the system watching for opportunities that could steer the framing.

40. 40

    05 Tips & tools

    Compact versions of the communicate habits: outline brainstorms instead of paper requests; reviewer personas that hold a view rigidly, run over drafts before real reviewers do; and a standing funder watch, with fit checked as the idea forms and aims steered as opportunities appear.

41. 41

    05 Brainstorm

    Facilitated brainstorm closing the loop on the survey theme. Everyone drew the same line about voice; here the room defines where that line sits in writing.

42. 42

    <> Close

    Section marker. Three small things to try this week, then thanks.

43. 43

    This week

    Three small moves that compound. Start the folder for one live project — messy is fine. Feed it one meeting transcript and ask for the synthesis you actually need. Prototype one output — the paper, dashboard, or site you’re picturing — rendered as a single file.

44. 44

    Thank you

    Closing invitation: reach out, bring a project, come experiment.

45. 45

    <> Appendix

    Appendix marker. Everything after this is reference material for people reading the deck afterwards — it was not presented live.

46. 46

    Glossary index

    Reference page for readers after the workshop: eleven terms, each with a full definition slide following.

47. 47

    Define · Prompt

    The most common confusion in a mixed-skill room. The prompt is the ask; the context is the material. Most disappointing answers are a context problem, not a phrasing problem.

48. 48

    Define · Token

    Why very long chats degrade, and why context limits exist. Tokens, not pages, are the unit.

49. 49

    Define · Harness

    Why the same model behaves so differently in different apps. Choose the harness for what it can reach — your files, your tools — not just the model name.

50. 50

    Define · Connector

    Connectors are how the context library stops being a manual upload chore. Also the moment to think about what data should and should not be reachable.

51. 51

    Define · Reasoning model

    Reasoning models spend more compute before answering. Use them for analysis and planning, not for everyday drafting.

52. 52

    Define · System prompt

    Custom instructions at the project level are the highest-leverage prompt work: written once, applied to everything.

53. 53

    Define · Multimodal

    Removes the excuse that material is not digital. Photos, scans, and recordings are all valid inputs.

54. 54

    Define · Hallucination

    Set up before the literature stop. The point is not fear; it is that verification is a designed step in the pipeline, not an afterthought. A confident tone carries no information about accuracy.

• The workshop — what was asked for, what faculty said they needed, and how the session was designed around it
• The methodology — the design-sprint method the pipeline is built on
• Health Design Sprint — the M4 elective referenced in the case study on slide 13