AI Content Platform
Content Authoring Assistant
An AI platform for building and revamping instructional content, component by component — verifying its own math and keeping a human in the loop.
Live demo — click around
reciprocal, quotient, dividend, divisor
Seeded from the diagnostic finding — “Add the term ‘inverse operation’.”
“Add “inverse operation” to the keyword list so it matches the standard’s language.”
Added “inverse operation” to the end of the list and kept the existing terms in order.
reciprocal, quotient, dividend, divisor, inverse operation
Interactive prototype with representative sample data.
The challenge
AI-generated instructional content is easy to draft but hard to trust: components reference each other, and a single wrong constant or unit can quietly propagate into claims and quiz answers. Authors needed a way to build content in the right dependency order, catch calculation and unit errors before publish, and see exactly how every lesson, claim, and question connects.
Our approach
The platform models each lesson as a knowledge graph linking lessons, standards, objectives, components, claims, questions, and verifier flags with explicit dependency edges. Authors build component-by-component in dependency order, and each component's context is injected from its prerequisites before Claude Sonnet (via Amazon Bedrock) generates a draft. A separate Python verifier sidecar checks calculations, units, and physical constants, flagging mismatches for human override rather than silently accepting or rejecting them. The same engine also powers a non-destructive lesson-revamp flow (shown in the demo above): it diagnoses an existing lesson against a target standard, then walks a curriculum lead through section-by-section revisions with the model — original and revised side by side — before writing only the approved changes back to the editor.
How it works
Build in dependency order
Authors add components one at a time following the graph's dependency edges, so each piece has its prerequisites in place before it is written.
Inject prerequisite context
When a component is generated, the platform assembles context from its upstream dependencies (prior knowledge, key terms, objectives) so the draft is grounded in what came before.
Generate drafts with Claude Sonnet
Claude Sonnet, running on Amazon Bedrock, produces the component draft. State, conversations, and prompts are persisted in S3.
Verify calculations and units
A Python FastAPI verifier sidecar independently checks the draft's calculations, units, and physical constants, and raises a flag on any mismatch instead of failing silently.
Human override on flags
Each flag surfaces the draft value against the verifier's expected value and requires a human to override, regenerate, or correct it before the content can move on.
Review and publish downstream
Content passes through a multiple-choice review gate, then publishes downstream via an SQS queue for delivery.
Also in the platform
The lesson knowledge graph
Beyond the revamp workflow, every lesson is modeled as a live graph — linking components, claims, questions, and verifier flags by dependency — so authors can see exactly how one change ripples through the rest. Hover a node to trace its connections.
Tech stack
Results
The system pairs LLM drafting with an independent verifier and a human-override gate, so calculation, unit, and physical-constant errors are caught and explicitly resolved before content publishes rather than shipping unnoticed. The knowledge graph keeps the relationships between lessons, components, claims, questions, and flags fully traceable for authors. Quantitative outcomes should be filled in from real production data.
Verifier flags caught before publish
Authoring time per lesson
Components authored / month
Override vs. regenerate rate on flags
Downstream publish throughput
Metrics to be populated with the project owner’s real figures.
Ready to Start Your AI Journey?
Let's build AI solutions that save time, reduce costs, and help your business grow. Book a consultation today.
