IDEEAS Lab

Adaptive AI Interviewer

A protocol-guided AI interviewer that helps research teams conduct adaptive semi-structured interviews while tracking topic coverage, conversational depth, and the remaining time budget.

Built on our work in faculty thinking: CAREER: Minds and Machines: Exploring Engineering Faculty Member Mental Models of Generative AI and Instructional Decisions.

Research prototype, internal, research testbed, actively developed.

The problem

Protocol-based interviews are difficult to scale without losing the judgment skilled interviewers use to balance required questions, productive follow-ups, unexpected insights, and limited time. Ordinary chatbots give researchers none of that: no reliable protocol coverage, no auditability, no study-management controls.

The system treats an interview as a coverage and information-acquisition problem. A locally hosted language model conducts the conversation while an orchestration engine decomposes responses into claims, matches them to protocol goals, tracks coverage and novelty, and uses a transparent policy layer to decide whether to probe, ask the next protocol question, change topics, or close. Participant consent and lifecycle controls sit beside researcher tools for protocol authoring, live monitoring, transcript review, and export.

For qualitative researchers, engineering education researchers, research teams.

What it looks like

Participant interview view showing the current protocol topic, remaining topic count, time remaining, and an adaptive follow-up question from the interviewer.
A participant interview in progress: the current protocol topic and time budget stay visible while the interviewer asks an adaptive follow-up. The exchange shown is seeded demonstration content.
Study entry page with a consent agreement, participant information form, and a study title labeling itself a UI-audit demo.
Consent gates the interview: participants see the study description, agree explicitly, and only then start. The study labels itself a demo in its own title.
Researcher study dashboard showing session counts, completion rate, duration and turn averages, a 30-day activity chart, and recent sessions with seeded participant names.
The researcher’s fieldwork view: sessions, completion, durations, turns, and live activity. The study title carries its UI-audit-demo label in-frame, and all participant names are seeded fixtures.
Session summary showing five AI-extracted key points organized under protocol topics, with session metadata and researcher notes beside them.
After an interview: AI-extracted findings organized by protocol topic, with the transcript and timeline one tab away. Every participant name in these researcher views is a seeded fixture; no real participant has used this platform.
Export screen offering JSON, CSV, or per-transcript ZIP formats, with separate opt-in controls for transcripts, summaries, and participant PII.
Research-data export with PII off by default: transcripts and AI summaries export cleanly toward downstream qualitative-analysis workflows, and identifying fields require an explicit opt-in.
Mobile participant interview view showing a participant answer about climate knowledge and the interviewer’s follow-up question.
The same interview, app-free on a phone. An earlier capture, so details may differ from the current interface; the conversation is demonstration content.

Research basis

The platform is the instrument side of the lab’s information-measurement research: if you can measure what an interview elicited, you can study whether an adaptive interviewer elicits more of it. Its faculty generative-AI interview protocol supports the lab’s NSF CAREER Minds and Machines project, and the system connects the lab’s qualitative methods, mental-model elicitation, and human-centered AI work in one experimental testbed.

What it does today

  • Conducts protocol-guided interviews with a locally hosted model, showing participants the current topic, coverage, and time remaining.
  • Decomposes responses into claims, matches them to protocol goals, and tracks coverage and novelty as the conversation moves.
  • Chooses the next interview action (probe, next question, topic change, or close) through a transparent policy layer.
  • Gives participants consent, pause, and withdrawal controls; gives researchers protocol versioning, study configuration, live monitoring, transcript and summary review, and exports with PII opt-in.
  • Carries a July 2026 verification record: 536 backend and 67 frontend tests, plus lint, type, build, migration, mobile, streaming, and export checks.

What it does not do

  • Not a public product: no public URL, no durable production deployment, and no way for outside researchers to run studies on it today.
  • No completed participant field pilot of this platform, and no documented external user feedback.

What it does not show yet

  • The interview policy is research-grade: coverage calibration, claim granularity, contradiction handling, and evaluator validity are open research problems, not solved ones.
  • A draft manuscript and pilot experiments exist but are unpublished; nothing here is a validated instrument.
  • Some measurement experiments use a public, participant-consented interview corpus; that is not evidence this platform has completed a field pilot.

Outcomes

  • Demonstrates an auditable end-to-end research system (participant safeguards, researcher workflows, model orchestration, and measurement infrastructure) rather than a chatbot demonstration.
  • Provides a controlled testbed for studying protocol adherence, information yield, saturation, and how AI interviewers influence what participants disclose.

Limitations and responsible use

  • The repository, study protocols, consent materials, and any future participant data stay private. Published captures use seeded demonstration content only.
  • Field use would require IRB approval and outcome-grounded evaluation of the interview policy first.

Evidence

Verified working prototype
A July 2026 verification record reports 536 non-live backend tests and 67 frontend tests passing, alongside lint, type checking, production build, migration, mobile, browser, streaming, and export checks. As of July 31, 2026.
Complete participant and researcher journeys
Study entry, consent, adaptive interview, coverage feedback, pause and withdrawal, and completion all run locally, as do researcher protocol, monitoring, review, and export workflows; twenty-nine UI-audit screenshots document the surfaces, six published here from seeded demo data. As of August 28, 2026.

Maintenance

Maintained by
IDEEAS Lab
Status
Research testbed, actively developed
Last reviewed
August 28, 2026

All tools