ADP

Learning Agent · Proof of Concept

CT101 “Time Fundamentals” · runs keyless on the mock provider · passcode gates removed for review

An AI learning tutor that knows exactly where the class is.

The Learning Agent is an AI tutor grounded in a course's own content — it answers from the ingested curriculum, never from the open internet, and never ahead of the room. In a live ILT session it follows the class and holds back topics the instructor hasn't reached (no spoilers); as a standalone tutor it opens the whole course with quizzes and ranked summaries. Around the tutor sits the instructor's live cockpit, engagement pushes, and analytics for the learning team — all demonstrated in this POC on one real course: CT101 — Time Fundamentals.

Grounded answers · no hallucinated content Coverage frontier · appendix never answerable Provider-portable: mock · Anthropic · Azure OpenAI · Bedrock Class-level sentiment · no learner identity
Featured · the transcription piece

🎙️ Live Transcription Demo — Script Sync

The POC isn't connected to Microsoft Teams, so it mocks a live course from within the POC itself: a simulated instructor "speaks" the CT101 verbatim control script as live captions — the same stream a Teams transcript feed would deliver — and Script Sync works out where the class is, in real time.

  • Watch it lock on — forward-only matching with anchor-phrase fingerprints and two-reading hysteresis (no wobble).
  • Make it misbehave — send the instructor off script (flags OFF SCRIPT after ~90 class-seconds) or into appendix material (out-of-course drift alert).
  • The frontier follows the class — the engine suggests a coverage move; the instructor confirms. Learners' gate updates live.
  • End class — get the after-class adherence report: segments taught, % on script, every detour timestamped.
▶ Open the transcription demo Upload an ILT script →
06:38Let's talk about two rules that people confuse constantly: rounding and gracing.
06:41Rounding snaps a punch to the nearest interval.
06:45Gracing looks at the schedule and asks: was this punch close enough…
ON SCRIPT · CT101-M3-08 “Rounding versus gracing” · segment 8 of 19
💡 Class appears to be on Rounding versus gracing — advance the coverage frontier to L3? [✓ Advance]
One agent · three experiences
🎓

Learner

In a live class, or self-serve afterwards
  • Tutor chat grounded in CT101 — concise, cited answers from the course content only.
  • No spoilers in-training — questions ahead of the class get a friendly hold (or a "jumping ahead" nudge with Show me anyway, if the instructor allows it).
  • Standalone mode — the full course, on-demand quizzes, and "summarize lesson 3"-style ranked takeaways.
  • Live engagement — instructor polls, word clouds, shared resources, and the end-of-session survey appear right in the chat.
  • 👍/👎 on every answer — feeds CSAT and the content team's quality views.
Open the tutor →
🎤

ILT Instructor

During the live session
  • Live cockpit — read the room: frustration, attention, deferred questions, CSAT, and a concept-struggle heatmap with a live nudge.
  • Drive the coverage gate — "Now on Lesson X", advance everyone, hard hold vs allow-jump-ahead.
  • Push to learners in one tap — pace check, confidence check, custom polls, word cloud, resources, end-of-session survey; responses stream in live.
  • Script Sync — see on-script / off-script, where the class is against the verbatim script, and pacing vs plan.
  • Presenter mode — a guided overlay that walks the demo on screen.
Live cockpit →Transcription demo →
📊

Trainer · L&D · Admin

Before and after the session
  • Dashboards — Instructor Insights, Module view (self-serve vs ILT, deflection, cohorts), Content & Quality (repeat rate, low-confidence log + CSV, quiz mastery), and a Hallucination Barometer.
  • Ingest any course — SCORM / Rise / Storyline upload, browser or headless deep-scan, video speech-to-text, or paste a transcript.
  • Ingestion pipeline (target state) — the six-stage demo: extract → chunk (structure from the manifest) → preflight battery (Gate A) → human review by exception (Gate B) → publish / version / rollback.
  • ILT Scripts — upload the learning team's verbatim control script; validated against the playbook rules on ingest.
  • Test console — switch LLM providers live, simulate the Teams feed, inspect the coverage matrix and retrieval explainability, seed demo data.
Dashboards →Ingest →Pipeline demo →Admin →
About this POC. Everything runs locally from one command with no API key (the default mock provider is deterministic and offline); switching to Anthropic, Azure OpenAI, or AWS Bedrock is a .env change — no code edits. The coverage rule ("is this chunk answerable right now?") is enforced at the retrieval layer, not in the prompt, and the appendix / supplemental material is never answerable in either mode — there are tests proving both. For review convenience every surface is open: passcode gates are removed, and mode/persona switches that would be driven by context in production (Teams meeting vs standalone app, SSO roles) are exposed as toggles. Ingested content, course expansions, and survey edits are in-memory overlays that reset on restart. See the README for architecture, deployment (Render / Docker / any cloud), and the full feature docs.