Your methodology, not the model's
A teaching assistant that lives inside the course, and generates only from the rubrics and exemplars your editorial team approved. Every paragraph carries a citation back to the source, so fidelity is something a reviewer can check rather than hope for.
8
stages from a teacher's ask to a cited lesson
1
source of methodology truth, owned by editorial
0
paragraphs generated outside the approved corpus
Five ways a published methodology erodes
A method that survives editorial review and then drifts in a thousand classrooms has not really been adopted — and generic AI accelerates the drift rather than slowing it.
Lesson prep eats the week
Building a single methodology-aligned lesson takes a teacher hours, and the ones built on Friday show it.
Methodology drifts in the wild
Once published, teachers remix lessons, and the publisher's method quietly erodes classroom by classroom.
Generic AI invents pedagogy
A general assistant produces plausible-but-off-method content: wrong grade band, wrong vocabulary, and sometimes the opposite of the publisher's position.
Co-pilots live outside the LMS
Switching tabs breaks the teacher's flow, and nothing generated in another tab carries provenance back to the source rubric.
No single source of methodology truth
Rubrics live in PDFs, exemplars in shared drives, and scope-and-sequence in spreadsheets, so an editorial update takes months to reach a classroom.
From a teacher's question to a cited lesson
The validation step before display is the one that matters: an output that fails the publisher's own rubric never reaches a classroom.
The teacher opens a course
The lesson is launched inside the LMS, and the integration hands over role, course and roster context.
LTI 1.3 Advantage
The co-pilot mounts in place
It loads inside the course the instructor is already building, not in another tab.
Iframe + JS API
Query understanding
An intent classifier resolves the ask — scaffold, rubric, exemplar, remediation — and extracts grade level and subject.
Structured outputs
Hybrid retrieval
Sparse and dense recall over the publisher's methodology base, filtered by grade band and methodology version.
Hybrid search
Rerank and assemble
Retrieved passages are scored by methodology fidelity, and the top ones assembled into a citation-tagged context window.
Reranking
Grounded generation
The lesson scaffold is generated constrained to that context, with a fast draft streaming in parallel for responsiveness.
Claude + fast model
Quality validation
Scored against the publisher rubric — fidelity, language, vocabulary, citations — before it is shown to the teacher.
Eval pipeline
Output with provenance
The lesson lands in the LMS with citations to the source rubric, and instructor edits flow into session history for editorial review.
Postgres + object storage
Hours returned to teaching
Building a methodology-aligned lesson went from a multi-hour task to roughly twelve minutes of a teacher's time.
That is what this engagement measured, not a promise about yours: the figure depends entirely on how much of your methodology is already written down. What transfers is the mechanism — the teacher edits a grounded draft that already cites its sources, instead of assembling one from a rubric PDF and a folder of exemplars.
Six capabilities, one embedded assistant
Methodology knowledge base
Rubrics, exemplars and scope-and-sequence ingested with grade-band and version metadata. Editorial approves what goes in, so the source of truth has an owner.
Hybrid retrieval and reranking
Keyword recall plus dense embeddings over your own knowledge base, then reranking to surface the highest-fidelity passages before anything is generated.
Grounded generation with provenance
Generation happens only from retrieved context, and every paragraph ships with a citation back to the rubric, exemplar or template it came from.
Embedded inside the LMS
Standards-based integration for the major platforms, and a custom embed where the platform has no standard. Teachers never leave the course they are building.
Low-latency drafting
A small fast model streams a first draft while the grounded, citation-carrying pass runs in parallel — so the page is never empty while the good answer is being made.
An eval pipeline on every release
A publisher-curated rubric scores fidelity, grade-appropriate language and citation correctness on every model or knowledge-base change. Below threshold, the release is blocked.
Wherever a method is the product
K-12 publishers
A writing-methodology publisher embedding the co-pilot inside adopting districts, so the method reaches the classroom as it was written.
Higher-ed adaptive learning
Scaffolding remediation content along mastery paths, with time-to-mastery instrumented against the platform's own cohorts.
Corporate training
A proprietary methodology rolled across a large workforce, with the co-pilot enforcing it rather than trusting each trainer to remember it.
Professional certification
Exam-aligned practice items drafted from the published body of knowledge, with citations that let a reviewer check the alignment.
Language-learning platforms
Generation grounded in a proprietary level-aligned methodology, so practice material stays faithful to the level it claims.
Editorial goes first
The knowledge base is built before anything is embedded, because a co-pilot with nothing authoritative to retrieve is just a chatbot in a course.
Weeks 1–4
Methodology ingestion
Editorial submits rubrics, exemplars, scope-and-sequence and templates. We chunk semantically, embed and store with version metadata.
Weeks 5–10
LMS integration and eval
Platform registration, embedding, single sign-on and rostering, plus the evaluation pipeline that gates every later change.
Weeks 8–12
Pilot cohort and tuning
A teacher pilot inside live courses. Retrieval, rerank weights and prompt scaffolds tuned against the disagreements the pilot surfaces.
Week 12+
Production rollout
Full rollout with usage analytics, version-pinned methodology, an immutable audit log and an editorial review queue.
Common questions
Asked by publishers, answered plainly
Every generation is grounded through hybrid retrieval over a methodology knowledge base your editorial team owns and approves. The model generates only from retrieved passages, and each paragraph carries a citation back to the rubric or exemplar behind it — so a reviewer can check fidelity rather than trusting it.
Canvas and Blackboard through LTI 1.3 Advantage, including names and roles provisioning and grade services; LearnWorlds through a custom iframe with its JavaScript API; and other platforms through their own REST interfaces where no standard exists.
Phase one is a four-week ingestion: editorial submits rubrics, exemplars, lesson templates and the adopted scope-and-sequence, which we chunk semantically and store with grade-band and version metadata. Updates are a re-ingestion, and the version is pinned, so a methodology release reaches classrooms in days rather than months.
A small fast model streams a first draft almost immediately while the grounded, citation-carrying pass runs in parallel. Teachers see something appear straight away and the higher-fidelity version replaces it, rather than watching a spinner while the good answer is assembled.
An evaluation pipeline runs against a publisher-curated rubric set on every model or knowledge-base change, scoring fidelity, grade-appropriate language, methodology vocabulary and citation correctness. A release that scores below threshold does not ship.
Yes, and permissions follow the role the LMS already knows: instructors get generation and remixing, administrators get methodology configuration and usage analytics, and editorial holds the queue that decides what enters the knowledge base.
Four weeks for methodology ingestion and the knowledge base, then around eight weeks for integration, the evaluation pipeline and rollout. The first four weeks depend mostly on how ready your editorial material is, which is the honest variable in this project.
Student personal information never leaves your LMS. The co-pilot operates on methodology content and instructor prompts, not student records, which keeps the sensitive surface out of the system entirely rather than protecting it after the fact.
Send us one rubric and one exemplar
This is our AI engineering practice
It is real work and it is where our four products came from. But what Cognilium leads with is narrower: optimization apps that run in tandem with Microsoft Dynamics 365, computing the decisions the ERP records but does not derive — the optimal price, the optimal pick path, the optimal stock level. See the optimization apps · How we build inside the ERP.