01 · Planner
Decomposes the source into a learning path
Parses the document or transcript, extracts topics and dependencies, and emits an ordered plan of modules and steps you can review, reorder, or rewrite before the session starts.
Multi-agent learning system
Four agents. One adaptive learning loop.
Repeater is not a single model behind a chat box. It is a pipeline of specialised agents that plans, teaches, tests, and profiles the learner so every next session adapts to you.
Planner · Explainer · Evaluator · Learner ProfilerPDFs, slides, transcripts, videoAdaptive sessions on every cycle
Source material
InputPDF12 pages · 1 transcript · 3 diagrams
Photosynthesis & Calvin cycle — notes.pdf
Lecture.mp4Transcript.txtSlides.pdf
Repeater pipeline
Adaptive sessionpipeline.ts
Guided learning loop
01 · Planner
Decomposes the source into a learning path
02 · Explainer
Teaches each step in the modality that fits
03 · Evaluator
Probes understanding with quizzes and open questions
04 · Learner Profiler
Builds the learner profile behind the next session
↻ loopstate.persist()→ next session
Architecture
A pipeline, not a chatbot.
Each agent owns one job and passes a typed contract to the next. Failures are recoverable, decisions are inspectable, and every module on the plan goes through the same loop.
02 · Explainer
Teaches each step in the modality that fits
Picks the right format per concept: short video, narrated walkthrough, mind map, flashcards, comparison cards, or animated diagram. Audio and visuals come from the same context, no copy-pasting between tools.
03 · Evaluator
Probes understanding with quizzes and open questions
Generates targeted multiple-choice items and open-ended prompts grounded in the step content, compares answers against the source, returns structured feedback, and re-explains the parts you missed so you can choose to dig deeper or move on.
04 · Learner Profiler
Builds the learner profile behind the next session
Aggregates signals from prior runs — wrong answers, skipped steps, hesitations, time-on-step — and updates a learner profile that synthesises the next adaptive session around weak topics and stable ones.
End-to-end flow
From raw source to adaptive session.
Every interaction is captured as state. The planner kicks off, the explainer and evaluator alternate per step, and the learner profiler closes the loop by shaping the next session.
01
Ingest
Drop in a PDF, slide deck, lecture video, or transcript. Repeater extracts text, audio, and structural cues to build a normalised representation.
02
Plan
The planner agent produces a typed learning plan — modules, steps, and step types (explain, quiz, summary). You can confirm it, edit modules, or regenerate before the session begins.
03
Teach
The explainer agent picks a modality per step: short narrated video, mind map, flashcards, animation, or comparison cards. Materials are rendered inline, not linked out.
04
Test
The evaluator agent issues quizzes and open questions. Responses are checked against the source and answered with structured feedback. Missed parts are re-explained — you decide whether to dig deeper or continue.
05
Adapt
Once a session ends, the learner profiler reads all signals, updates the learner profile, weights gaps by recency and severity, and produces a new adaptive session that targets what is still weak.
Multimodal output
Visual, auditory, kinesthetic — chosen per step.
The explainer agent does not default to walls of text. It selects the modality that fits the concept and the learner's prior interactions.
Narrated video
Short clips with synchronised voice-over and on-screen visuals, generated from the source and the planned step.
Mind maps
Concept graphs that expose the topology of a topic before drilling into the details.
Flashcards
Atomic prompt/response pairs scoped to the current step, surfaced exactly when the evaluator decides recall is the right test.
Animations
Step-by-step animated diagrams for processes, cycles, and mechanisms where motion carries the meaning.
Adaptive sessions
The learner profile is what makes the loop adapt.
After each session, the learner profiler compiles an interaction trace — answers, confidence, time-on-step, retries — updates the learner profile, and writes a new session plan that re-balances coverage. The longer you use it, the more precisely it converges on your weak spots.
Per-learner state
Every session writes to a persistent learner state. Future sessions read from it; nothing is regenerated from scratch.
Targeted re-teaching
Topics with failed evaluations come back first, with a different modality and a different angle, not the same explanation twice.
Editable plan
Every module on the plan shows what it will cover before the session starts. Rename, reorder, remove, or regenerate any step — the loop runs on your version, not a black box.
Where a multi-agent loop beats a single chat.
Anywhere the work is more than answering one question — where you need a path, a check, and a way back when something does not stick.
University study
Turn a 60-page chapter into a plan you can finish in an evening, with quizzes that reveal what you actually retained.
Professional certifications
Long syllabi broken into testable modules. The learner profile concentrates time on the areas you keep failing.
Internal training
Onboarding documents become guided sessions with evaluation built in, instead of PDFs nobody reads end to end.
Self-directed learning
Drop a paper or a long video and let the planner expose the structure before you commit hours of reading.
Technical questions
Why a multi-agent system instead of one large prompt?+
Each agent has a narrow contract — the planner returns a typed plan, the evaluator returns structured feedback on the answer — so failures are localised and the pipeline is inspectable. One large prompt would conflate planning, teaching, and feedback, and degrade on long documents.
Can I edit the plan before the session starts?+
Yes. The planner agent emits a learning plan with modules and steps. You can rename, reorder, delete, or regenerate any module before entering the teach/test loop.
What happens when I get an answer wrong?+
The evaluator does not grade you and does not gate progression. It returns feedback and re-explains the part you missed. From there you decide whether to dig deeper on the topic or move on with the rest of the session.
What does the learner profiler actually do?+
It reads the interaction trace from prior sessions — wrong answers, retries, time-on-step, skipped steps — updates a learner profile, and emits a new session plan that re-prioritises weak topics with different modalities.
What sources work as input?+
PDFs, slide decks, lecture videos, audio recordings, transcripts, and free-form prompts. Audio and video are transcribed and aligned before the planner runs.
Run the loop on your next document.
Drop a source, watch the planner produce the path, then move through teach and test until the learner profiler has enough signals to adapt the next session.
Start a Repeater session