fkra AI · configurable agents

Build the agent. Hand it the tools.

Every stage of learning runs on an agent you configure: its model, its tools, its guardrails. Arm a tool and the agent’s work changes with it. Nothing is hard-wired.

Agent builderlive
RafiqCourse Builder

building module 9 / 12

On shift16 of 30+ · scroll

Course Builder · 3 of 8 tools armed

Modelclick to swap
Tools3 of 8 armed
Run traceRafiq · Course Builder600ms
retrieveretrieve(policy_v4) → 12 chunksquizgenquiz.write(module_3) → 6 itemsdiffdiff(v3 → v4) → 9 changesoutputcourse blueprint

Click a tool to arm or disarm it; the trace re-runs live

ParseTranscribeBlueprintCiteOCRQuizRehearseDiarizeGradeTranslateVersionAudit
01 · The pipeline

Five agents, five jobs. Try each one.

Each stage below is live. Queue a source, pick a clause, cut a module, hold the gate, run the build.

Median import → approved draft: 6 min
Intake agent

Queue anything. Lineage sticks.

Every source gets an id the moment it lands. Six stages later, a quiz answer still points back to the file it came from.

Click a source to queue or remove it

Ingest queue2 queued
src_01 · pricing-policy-v4.pdfreadsrc_03 · incident-drill.mp4read

Each id survives every downstream step; a quiz answer still points at pricing-policy-v4.pdf.

PDFpricing-policy-v4.pdf · p.7Scanning
Parse agent

Reads it like a lawyer.

Clauses, tables, and exceptions come out anchored, so nothing floats free of its source.

ExtractedAnnual contracts signed after Mar 1 extend to 30 days.
§3.2aexception Quiz item · scenario

Pick a clause; it holds while you read

Curriculum agent · 03

Structure before a word of prose.

The plan arrives as modules you can cut. Click one to drop it; the run time and quiz count follow.

5modules45mrun time14quiz items
Human gate

No agent publishes itself.

A named owner signs the blueprint before a single lesson is written. Every decision is logged against their name.

You hold this one. Try it.

Reem · L&D leadOwner of this blueprintAwaiting sign-off
scope5 modules · 45 min · 14 quiz itemssources2 sources, all lineage resolvedpass barPass bar 80% · retries 2
Generation is blocked until this clears
Generation agents · 05

Modules land with receipts.

Writer, quiz, and fact-check agents work in parallel. Every claim carries the clause it came from.

Build not started0%0 citations attached
M1 · Why pricing rules existRafiq · Course Builder · Claude● buildingwriting…
M2 · Refund windows in practiceRafiq · Course Builder · Claude○ queued
Quiz · 14 itemsLina · Skills Mapper · GPT○ queued
Rehearsal · angry renewalSara · Learner Tutor · Claude○ queued
Fact-check passOmar · Compliance Auditor · Gemini○ queued
02 · Multi-modal input

Feed it anything with a lesson inside.

Policies live in PDFs. Expertise lives in recordings, decks, diagrams, and half-watched YouTube links. fkra reads all of it. Pick a source type.

One pipeline, six input types
VIDEOincident-drill.mp4parsed
Transcript, diarized42:108 chapters cutby topic shift5 quiz moments pinnedto timestamps
Becomes Chaptered lesson Timed quizTimestamps are citations
03 · Four learning modes

One module. Four ways in.

See it, hear it, read it, do it. Every module is generated in all four at once, so Universal Design for Learning is the default output rather than a retrofit. Pick a mode to see what the learner gets.

4 modes per module, alwaysLearner picks, or fkra picks for them
Visual modeM2 · Refund windows in practice9 min

A decision tree, not a wall of text.

The refund rule is drawn as a branching path with the clause pinned to each branch. Color carries no meaning on its own.

Annotated decision tree4 branchesThreshold table, color-safeT-2Clause pinned per branch§3.2 · §3.2a
AccessibilityAlt-text per nodeColor-blind safeZoomable SVG
UDL principle servedMultiple means of representation

Learners who process visually get the same governed content, not a simplified version of it.

Representationsee · hear · readAction & expressiondoEngagementlearner choice
Per module, every time

4modes generated from the same cited source, no extra authoring pass.

04 · Governed by design

Every output carries its source.

Generative AI in a compliance tool only works if every output can be checked. Every module, quiz, and answer here carries the clause it came from.

Citations first

Every generated line points back at your document.

Fact-check pass

Mixed claims get flagged before review, not after launch.

Human gate

Nothing ships without a named approver.

Sealed tenants

Your documents train nothing. Retrieval only, isolated per workspace.

Early accessAI-assisted gradingRubric-based first-pass grading for written answers, with the reviewer always in charge. Rolling out to early-access teams.Join the waitlist
05 · Multi-provider

Bring your own model.

fkra is model-agnostic by contract. Route each stage to the provider you approve; click any chip to swap.

Zero retention · SOC 2 · per-workspace routing
Routing · Nimbus Retail workspace
Parseread + extract
Blueprintplan the course
Generatewrite modules
Fact-checkverify vs source
Translate10 languages

Click a provider chip to swap it. Routing is per stage, per workspace, and audit-logged.

Claude · Anthropiclong contextGPT · OpenAIfast draftsGemini · Googlevideo nativeMistralEU residencyLlama · self-hostedyour VPC
Your cloudOr run it in your VPC.

Open-weights models, self-hosted. Documents and prompts never leave your network.

6input types · four modes each
34citations per course, median
30+specialist agents on shift
100%versions human-approved
العربيةEnglishहिन्दीاردوবাংলাFilipinoBahasa IndonesiaFrançaisTürkçe中文

Author once, deliver in ten. Arabic is native, laid out right to left rather than mirrored with CSS.

Your next step

Put your documents to work.

Import one policy. Get a governed course, with receipts, in minutes.

Explore more