Studio

Self-Regulated Teaching Intelligence

An instructor-facing research prototype that turns ordinary class recordings into inspectable, evidence-linked feedback for post-class reflection.

Explore itResearch prototype · Internal · Live-classroom pilot ahead
SRTI live analysis view showing session metadata, a four-stream pedagogical timeline, and the speaker-attributed transcript.
ScreenshotThe live analysis view: session metadata, the four-stream pedagogical timeline, and the transcript in one place. The session is a public MIT OpenCourseWare lecture analyzed by the live pipeline — not private classroom data.

The problem

Instructors get little timely, structured evidence about their own classrooms: end-of-semester evaluations arrive too late, peer observations are rare, and memory is a weak basis for judging talk balance, wait time, questioning, or participation.

The approach

Class audio becomes a speaker-attributed transcript and four streams of pedagogical events — content, instructor moves, student engagement, and reflection triggers — alongside deterministic measures like talk ratio, wait time, and question rate. The pipeline is deterministic-first: the language model proposes events, deterministic code decides, and every event is verbatim-anchored so an instructor can jump from any claim to the passage behind it.

Research basis

Self-regulated teaching is the instructional counterpart to self-regulated learning: instructors need evidence to set goals, monitor their practice, and adapt. SRTI turns that construct into a working instrument, and its pilot — 20 open-courseware lectures, human-ratified reference labels, per-stream accuracy results — is the lab studying its own tool before trusting it.

What it does today

  • Ingests class audio, transcribes it locally, and separates speaker turns with editable attribution.
  • Organizes transcript-grounded observations into four streams: content events, instructor moves, student engagement, and reflection triggers.
  • Computes deterministic measures — instructor/student talk ratio, wait time, turn counts, longest monologue, question rate and type.
  • Lets an instructor filter the session timeline by stream or speaker and jump from any claim to its supporting passage.
  • Keeps analysis version history, with model and schema metadata on every run.
  • Ships a reproducible evaluation harness with a 20-lecture human-ratified reference corpus.

What it does not do

  • No hosted product, self-service onboarding, or production support — a private, internally deployed prototype.
  • Not validated in live classrooms; the pilot corpus is public courseware, which underrepresents student interaction.
  • Audio-only: it cannot observe body language, movement, visual materials, written work, or silent engagement.

Inside the prototype

A selected student contribution linked to the supporting transcript passage.
ScreenshotEvery analytic claim links to the transcript passage that supports it — here, a student contribution and its source. Public courseware audio; speakers appear only as anonymous labels.
Reflection-trigger cards with urgency labels and transcript evidence.
ScreenshotReflection triggers with urgency labels and transcript evidence — the stream that feeds post-class reflection, and the weakest performer in the pilot (F1 0.22 after rubric revision). Labeled experimental for exactly that reason.
Architecture diagram: class recording through ASR and diarization, deterministic-first LLM extraction, validation, metrics, and dashboard feedback.
DiagramThe deterministic-first pipeline: the language model proposes events per chunk, mechanical code merges and validates, and every event is verbatim-anchored before it reaches the dashboard.
Timeline figure showing 91 verbatim-anchored pedagogical events across four streams during an 80-minute session.
Diagram91 verbatim-anchored events across the four streams in one 80-minute session — live platform extraction from a public MIT OpenCourseWare lecture.

Evidence

Working pipeline on real lecture audio

The screenshots are live pipeline output over an 80-minute public MIT OpenCourseWare lecture — 91 verbatim-anchored events across four streams — not seeded fixture data.

As of 2026-08-28

Measured, including the failures

A 20-lecture pilot across five disciplines: 1,346 candidate events, 1,112 retained after human ratification, 100% verbatim-anchor resolution, and per-stream accuracy reported honestly from F1 0.68 down to 0.22.

As of 2026-07

Engineering depth

1,363 automated tests documented in a dated internal evidence memo, alongside a reproducible evaluation harness.

As of 2026-07-06

What we are not claiming

  • The pilot shows feasibility, not that the system accurately evaluates teaching. Concrete student-engagement events were strongest (detection F1 ≈ 0.68, 0.86 type agreement on matched events); instructor moves reached only ≈ 0.33 and reflection triggers 0.22 after rubric revision. The weak streams are named on purpose.
  • The pilot is small and bounded: 20 lectures, one run each, no confidence intervals, and reference labels that were LLM-assisted then human-ratified — not an independent expert-coded gold standard.
  • It is a professional-development and research tool. It is not designed, and should not be used, for employee evaluation or student surveillance.

Responsible use, privacy, and rights

  • Real classroom audio, transcripts, student identities, speaker maps, and instructor analytics are never published, and live-classroom use requires consent, IRB, and privacy review.
  • Interface imagery uses public MIT OpenCourseWare material with attribution; no private classroom data appears.

More from the Studio

All entries

Use it

Course Sphynx

A faculty-facing platform for helping instructors redesign courses through guided, transformation-oriented support.

Explore it

GATOS qualitative analysis workflow

A research workflow for using open-source generative AI and machine learning to support inductive qualitative codebook development.

Explore it

CATS QDA

A local-first workspace where qualitative researchers analyze sensitive interview data with AI assistance they can inspect, override, and audit.