The agent-native learning platform

Every lesson begins with a real job.

PAUL turns live job descriptions into complete, learning-science-backed curricula — generated by AI agents, taught through Socratic coaching, and proven in a record the learner owns.

Canvas LTI 1.3 & 1.1 · OAuth 2.1 · Built for MCP agents

P ersonalized

A gentic

U nified

L earning

One hiring need, start to finish

01The job description

Instruction starts where employment starts.

Nothing in PAUL is taught because it's in chapter seven. Every tutorial and concept lesson traces back — structurally, in the data model — to a competency profile built from a job an employer actually needs to fill.

That isn't a content policy; it's the architecture. Generating instruction requires a real profile and a concrete client scenario before a single word is produced.

Job posting · Meridian Health Systems

Junior Full-Stack Developer

Hiring now

We're looking for a developer to join our patient-portal team. You'll design and maintain RESTful APIs backed by PostgreSQL, and ship features through our CI/CD pipeline.

Day to day you'll build accessible UI in a modern component framework, harden authentication flows for HIPAA-regulated data, and write automated tests alongside the code they protect.

parse_job_description() · 6 employable skills detected

A real posting is the only valid starting point — the generation tools refuse anything else.

02The skills analysis

One posting, six labor-market taxonomies.

An employer's agent parses the posting, then cross-references every skill against NICE 2024, O*NET, ESCO, LinkedIn, Indeed, and Credential Engine — merging cross-source duplicates semantically into one harmonized graph.

The output is the honest list of employable skills a learner needs to qualify for an interview — grounded in labor-market data, not a committee's best guess.

Harmonized skills graph

"Design web service APIs" O*NET "REST API development" NICE
merged · cosine similarity 0.94
RESTful API design NICE O*NET
  • PostgreSQL schema design O*NETESCOLinkedIn
  • CI/CD pipeline operation NICEIndeed
  • Automated testing ESCOO*NETCred. Engine
  • Secure authentication NICEESCO

03The curriculum plan

Instructors scope the role, honestly.

From a published profile, an instructor chooses exactly which concepts a course will cover — a focused slice or the whole role — and PAUL computes the coverage as a percentage of the job, not a page count.

Guardrails flag over-promising: a profile too big for one semester gets split into a course sequence instead of being crammed into it.

Curriculum plan · CS 3420

Full-Stack Web Fundamentals

Scoped from: Junior Full-Stack Developer profile

In scope this course

  • RESTful API design
  • PostgreSQL schema design
  • Secure authentication
  • Automated testing

Deferred

  • CI/CD pipeline operation next course
  • Container orchestration next course

Guardrail: plans above a 50% soft ceiling are flagged for a course split — one semester shouldn't pretend to cover a whole job.

04Time-boxed sprints

Built to fit a real semester.

The plan is chunked into time-boxed sprints — sixteen weeks, seven two-week sprints, and buffer by default, with every parameter adjustable to your academic calendar.

Each sprint is anchored to a client story, so learners always work inside a believable business scenario rather than a numbered exercise set.

Semester map

16 weeks · 6 hrs/week · ~84 hours

Sprint 3 · weeks 5–6 · client story

“Meridian's patient portal locks users out after password resets — ship a corrected, secure authentication flow.”

Every sprint carries a required scenario like this one. It becomes the setting for the sprint's tutorial.

Weeks, hours per week, sprint length, and buffer are all parameters — quarters, 8-week blocks, and bootcamp cadences fit the same math.

05The tutorials

Watch an instructor's agent build the course.

Instructors work from their own AI agent — Claude, Cursor, any MCP client — driving PAUL's custom tools. A durable workflow engine generates each tutorial step by step and retries transient failures automatically, so long-running generation simply finishes.

instructor's agent — any MCP client

concept-mcp-server

Generate the Sprint 3 tutorial from our curriculum plan.

generate_sprint_tutorial({ curriculum_plan_id: "cp_8f2a…", sprint_number: 3 })

durable workflow started · run wf_c91d…

hydrateInputs

findSimilarTutorials

generateTutorialJSON 1 retry — transient timeout, auto-recovered

validateAgainstSchema

persistTutorial

indexForContextAlignment

Published: “Fixing Meridian’s Password-Reset Lockout”

9 steps · 6 concepts tagged · ZPD band: within

06Concept mining

Every tutorial step knows what it teaches.

As tutorials are generated, agentic tools tag the concepts inside every step. Full concept lessons are generated lazily — the moment a learner needs one, in the same business context they're already working in.

And when a learner struggles, remediation lanes generate easier stepping-stone tutorials and concepts — instead of replaying the same explanation, louder.

Tutorial · step 4 of 9

Wire up the JWT refresh flow

Replace the expiring session check with a refresh-token exchange so portal users stay signed in through a password reset…

concepts_in_step — mined from this step:

Token-based authentication assessedHTTP middleware assessedSecure cookie storage

Remediation lane · generated only if needed

Struggling with token auth? PAUL generates a stepping-stone lesson — “The HTTP request/response cycle” — routed by Gagné's intellectual-skills hierarchy, then brings the learner back.

07 · The engine underneath

Learning science, running as code.

Most platforms cite the research. PAUL executes it: five instructional-design frameworks are encoded as schemas, prompts, and validators inside the generation pipeline — and an automated quality score rejects any lesson that doesn't meet the bar before it ships.

Concept Coach · Socratic checkpoint

Coach: Your refresh token just leaked. What breaks first — and why doesn't the attacker get access forever?

Learner: They can mint new access tokens… but rotation means the next legitimate refresh invalidates the stolen one?

Coach: Close. Walk me through what the server sees when both tokens try to refresh.

Mastery is demonstrated in dialog — never multiple-choice guessing.

  1. Vygotsky — Zone of Proximal Development

    Tutorials open fully scaffolded and deliberately fade support, targeting learner independence about 70% of the way through.

    zpd_metadata.target_independence_step
  2. Gagné — Nine Events of Instruction

    Every lesson is scored against Gagné’s nine events; remediation routes down his intellectual-skills hierarchy.

    gagne_events ≥ 6/9 required
  3. Merrill & Tennyson — Concept Teaching

    Concepts are taught through critical attributes with matched examples and near-miss non-examples, never definition dumps.

    positive/negative example balance 40–70%
  4. Merrill — First Principles of Instruction

    Lessons move through activation, demonstration, application, and integration around real problems.

    fpi_phase: activation → integration
  5. Socratic Coaches + CARL Reflection

    The Practice Coach and Concept Coach probe understanding through dialog, closing with Context-Action-Result-Learning reflection.

    carl_reflection · gagne_event 9

08The record

Proof of understanding the learner owns.

Every demonstration of mastery is captured per concept — recognition, explanation, application — and linked to the evidence behind it. Learners keep it as a Learning & Employment Record in a personal wallet, shared with employers through revocable, time-bounded links.

Learning & Employment Record · credential

Token-based authentication

Recognition

Explanation

Application

  • Sprint 3 tutorial — auth flow shipped verified
  • Concept Coach transcript — Socratic checkpoint verified
  • GitHub portfolio — prior work provisional

Share link · expires in 30 days · revocable anytime

paul.…/w/9tk2

Evidence earned inside PAUL is verified; outside evidence stays provisional until demonstrated — the distinction employers actually need.

PAUL

Personalized Agentic Unified Learning — from a real job description to a record of understanding the learner owns.

© 2026 PAUL · EdTech4Learning

Canvas LTI 1.3 & 1.1 · OAuth 2.1 · MCP-native