Technical strategies
Caffeinated Professor
The course-bounded mimetic agent is grounded in official course materials (lectures, slides, textbook, rubrics) and provides verifiable citations where possible (e.g., lecture timestamp, slide title, chapter/section). To increase reliability, the agent displays cascading fidelity: answers that cannot be derived directly from course materials are clearly differentiated (Distinct text color/font. Distinct voice tone.)
-
Source-grounded retrieval (RAG) over course materials.
Training content indexed as searchable chunks (lecture transcripts aligned to timestamps, slides with headings,
and textbook sections). Queries retrieve the most relevant passages before answer generation.
-
Citations/timestamps where possible in responses.
Responses include a short “Sources” block with 1 to 3 pointers to the exact material used
(e.g., “Lecture 3 @ 12:40–14:05”, “Slides: Stakeholder Theory—Objections”, “Textbook Ch. 5 §2”).
Citations use stable IDs to remain valid across transcript re-exports and content updates
(e.g.,
L03_T12m40s, SLD_Stakeholder_Objections, TXT_BE_CH5_S2).
-
If system cannot locate strong sources, clarifying questions are asked before relying on general knowledge.
-
Attribution integrity checks (anti-hallucination).
Citations validated against indexed corpus. System not allowed to invent timestamps, slide titles,
or readings. When retrieval confidence is low, it cascades down to “limited answer” mode.
-
Mimetic style layer.
A constrained persona prompt (+ a library of Q&A exchanges) enforce professor’s characteristic vocabulary and teaching, while preventing drift into a generic tutor voice.
-
Privacy-by-design deployment.
Data-minimization beyond that necessary for learning analytics and improvement; configurable data-retention policies suitable for institutional environments.
Caffeinated Professor assessment engine
After posing a question and receiving a response, the distance between the student response and a proficient answer is calculated. To reduce the distance, the agent provides cues and nudges for a subsequent response. The cycle continues until reaching the standard. The resulting grade derives from the amount of help required to reach proficiency.
This assessment requires a distancing calculation (between student and proficient responses) and scaffolding support (leading the student toward proficiency).
-
Distance grading compares a proficient response with the student response. The process splits into parallel tracks—one for the proficiency standard and one for the student—before meeting in a comparison engine.
-
Expert path (Generating the standard)
Trigger: Question submitted by human professor or generated from a source-grounded retrieval (RAG) over course materials.
Proficiency response: Agent predicts an answer corresponding with course materials, or human professor provides a gold standard answer.
-
Student path
Oral/Text Input and Capture: Student provides their response (text or verbal), which is prepared for analysis.
-
Comparison engine (Measuring distance)
Vectorization: Proficient answer and the student answer converted into mathematical vectors.
Distance calculation: The distance (similarity) between vectors is measured in semantic space or content space.
Attribution integrity check: Engine validates the student's answer against the indexed corpus to ensure they aren't "hallucinating" facts that aren't in the course materials.
Risk: System mostly measures semantic style/verbosity, so a correct answer phrased differently looks “far” while a fluent wrong answer looks “close.”
-
Scaffolding responds to the distance between the proficient response and the student's response by leading the student through discussion toward proficiency.
- Nudge: The agent provides a clarifying question or a targeted hint based on a specific slide or textbook section.
Tutor Mode required to forbid the agent from providing the standard answer. It should only recite from source-grounded retrieval (RAG) fragments marked with a specific lecture timestamp or slide heading.
-
Grading is a function of the quantity of hints required to reach the gold standard answer.
- Logging the Dialogue: Every nudge provided by the agent is logged, and final grade derives from the count of logs in the session.
Anti-Hallucination: Hints validated against the indexed corpus so they don't lead toward incorrect or "generic" answers.
-
Privacy-by-design deployment.
Data-minimization beyond that necessary for learning analytics and improvement; configurable data-retention policies suitable for institutional environments.