Browse freely. Only an explicit product selection changes your total. Nothing you already own is added automatically.
Core → every specialization: Core purchase required. All seven paths follow the same rule.
Foundation
AI Production Systems Core
US$49712 months · one-time
Start making useful decisions with AI, even if you have never coded or used an AI tool. Practise choosing the right task, giving clear instructions, and checking what comes back.
What you can practise now
✓Repair a flawed brief, save a reusable task instruction and check a changed practice case.
✓Build source-linked comparison briefs, preserve unresolved questions and reuse your recipe on different evidence.
✓Decide when AI fits a task and when a simpler approach is better.
37 current labs + 1 optional · Feedback as you practise
A current exercise: Get a useful first result
Inspect included labs +
Get a useful first result
Should this task use AI at all?
What is an unsupported model claim?
How does a language model produce the next token?
What is an application-supplied system prompt?
How do you build a robust prompt contract?
What does data quality mean?
Where should an AI system obtain information?
Why AI Splits Your Text Into Tokens—and What That Changes
What is a context window?
How should an AI request be packed?
How do you verify a model response claim by claim?
Write a brief you can stand behind
Why can evidence inside the context window still fail?
When should an AI answer, abstain, or escalate?
How should sensitive data flow through an AI system?
What can a system prompt not enforce or standardize?
Keep your assistant inside its boundaries
Build a document assistant you can actually use
Make AI output dependable
What is an AI agent?
How does a model use a tool?
Where did tool use fail?
Turn a batch into reviewed records
How do you choose evaluation methods?
How do you test whether a prompt change helped?
What does a benchmark score mean?
What does model size mean?
How should you choose a model size?
What does one model request cost?
What does a successful AI task really cost?
Find the first fault
Which layer owns this AI-system failure?
Release a workflow someone else can use
Turn unfamiliar evidence into a decision
Repair the workflow. Keep the evidence.
Make the method work on a new task
What changed because of the system prompt? Optional practice
Assessment & output: Save reusable instructions, checked briefs and workflows; later tasks build a document assistant, release evidence and portable work. An assessed credential is available for the named saved lab tasks—not for purchase or attendance. See credential scope.
Your starting point for every specialization. You do not need coding or AI experience to begin.
Learn what changes when AI can take actions. Practise deciding what an agent may do, what it should remember, and how to recover when an action goes wrong.
What you can practise now
✓Compare one model call, a fixed workflow and agent-directed control; choose the simplest option that meets the task and bound its authority.
✓Set tool boundaries, delegated authority and durable state in bounded exercises.
✓Inspect skills and MCP connections; compare multi-agent and computer-use choices.
16 current labs · Feedback as you practise
A current exercise: Choose a fixed workflow or agent-directed control
Inspect included labs +
Choose a fixed workflow or agent-directed control
What makes a tool-using model an operational agent?
What authority should an agent receive?
How do you prevent duplicate or uncertain agent actions?
How should AI memory stay fresh, attributable, and revocable?
What is an AI skill?
How should skills be described, selected, and loaded?
When is a skill package safe and appropriate to run?
What is the Model Context Protocol?
How does an MCP exchange work?
A valid MCP connection is not permission.
How should an MCP host handle timeouts and uncertain side effects?
When should multiple agents coordinate, and which pattern fits?
When should an agent control a computer interface?
How should an agent incident be contained and safely restarted?
Agent Operations Runbook
Assessment & output: Work through the current scenarios and inspect feedback on your decisions. An assessed credential is available for the named saved lab tasks—not for purchase or attendance. See credential scope.
Before you buy
Further Agent labs, an upgraded runbook assessment and an integrated capstone are still being developed. The control-choice lab and every exercise listed above are available now.
No date is promised for future additions. Choose based on what you can use today.
This path requires valid Core purchase history. For a first purchase, selecting a path also selects Core: US$994 total for Core + this path. You can buy Core now and add a path later.
Explore the building blocks behind reusable AI coding assistance. The current starting set focuses on skill packages: what they contain, when to use them, and whether to trust them.
What you can practise now
✓Inspect a reusable skill package and the instructions it carries.
✓Practise routing a task to an appropriate skill.
✓Identify trust and permission boundaries before using a skill package.
3 current labs · Feedback as you practise
A current exercise: What is an AI skill?
Inspect included labs +
What is an AI skill?
How should skills be described, selected, and loaded?
When is a skill package safe and appropriate to run?
Assessment & output: Work through the current scenarios and inspect feedback on your decisions. No credential is currently issued for this specialization. Core and Agent credentials are separate. See credential scope.
Before you buy
An end-to-end Vibe Coding, software testing and delivery curriculum is not yet available. These three labs do not constitute that full workflow.
No date is promised for future additions. Choose based on what you can use today.
This path requires valid Core purchase history. For a first purchase, selecting a path also selects Core: US$994 total for Core + this path. You can buy Core now and add a path later.
Practise spotting unsafe authority and checking safeguards in AI systems. Start with bounded decisions about tools, skills and recovery; add optional incident-leadership practice when useful.
What you can practise now
✓Constrain agent authority and examine the trust boundaries of MCP and skill packages.
✓Test whether an operating control addresses the risk it is meant to contain.
✓Practise incident recovery, with optional crisis communication and recovery decisions.
6 current labs + 3 optional · Feedback as you practise
A current exercise: What authority should an agent receive?
Inspect included labs +
What authority should an agent receive?
What is the Model Context Protocol?
When is a skill package safe and appropriate to run?
A valid MCP connection is not permission.
Does the control actually operate?
How should an agent incident be contained and safely restarted?
How should an incident be contained under uncertainty? Optional practice
What should each stakeholder be told, and when? Optional practice
When is an incident actually resolved? Optional practice
Assessment & output: Work through the current scenarios and inspect feedback on your decisions. No credential is currently issued for this specialization. Core and Agent credentials are separate. See credential scope.
Before you buy
The full attack-lab sequence, red-team assessment and proposed final project are not yet available.
No date is promised for future additions. Choose based on what you can use today.
This path requires valid Core purchase history. For a first purchase, selecting a path also selects Core: US$994 total for Core + this path. You can buy Core now and add a path later.
Learn to question an impressive score and build more useful tests. Practise finding leakage, checking evaluators and matching a decision threshold to the cost of being wrong.
What you can practise now
✓Choose data splits and identify misleading benchmark claims or evaluation leakage.
✓Check evaluator reliability and practise calibration and decision thresholds.
✓Assemble and challenge evaluation, assurance and release evidence in the existing exercises.
15 current labs · Feedback as you practise
A current exercise: When do benchmark results justify a product claim?
Inspect included labs +
When do benchmark results justify a product claim?
What are training, validation, and test sets for?
How should a dataset be split?
How do leakage and test-set reuse corrupt evaluation?
Can this evaluator or LLM judge be trusted?
What does a preference rating mean?
Are the model’s confidence estimates calibrated?
How should calibrated probabilities drive actions?
What does the balance laboratory test?
Why can evaluation evidence stop being useful?
Is the evidence strong enough for a product decision?
How do you build a safety assurance case?
What defeats or invalidates a safety assurance case?
Is the evidence strong enough to release?
Model Evidence Pack
Assessment & output: Work through the current scenarios and inspect feedback on your decisions. No credential is currently issued for this specialization. Core and Agent credentials are separate. See credential scope.
Before you buy
The proposed new measurement labs and upgraded evidence-pack assessments are not yet available. Existing exercises retain their current scope.
No date is promised for future additions. Choose based on what you can use today.
This path requires valid Core purchase history. For a first purchase, selecting a path also selects Core: US$994 total for Core + this path. You can buy Core now and add a path later.
Practise decisions that make an AI feature more useful and dependable: finding relevant information, handling memory, controlling failures, and balancing serving cost with capacity.
What you can practise now
✓Compare embeddings and retrieval choices; diagnose prompt and information-access failures.
✓Practise memory freshness, model specialization, tool reliability and release decisions.
✓Evaluate serving capacity, economics and customer lifecycle choices; explore optional architecture and training depth.
16 current labs + 11 optional · Feedback as you practise
A current exercise: What is an embedding?
Inspect included labs +
What is an embedding?
Is this embedding model fit for the retrieval task?
What is a vector database?
How should vector retrieval be tuned?
When is prompting the wrong lever?
How do you keep vector retrieval authorized, fresh, and deletable?
How should AI memory stay fresh, attributable, and revocable?
What changes when a model is specialized?
How do you validate a specialization safely?
How do you prevent duplicate or uncertain agent actions?
What limits serving throughput and latency?
Should this serving schedule be approved?
How can a prompt use fewer tokens without losing its contract?
How does the customer lifecycle work?
Where is the customer lifecycle constrained?
Which lifecycle intervention should you fund?
How should a training-data mix be composed? Optional practice
What is training loss? Optional practice
What should you do when training curves diverge? Optional practice
When does preference optimization go wrong? Optional practice
What is model-generated training data? Optional practice
When should synthetic-data generation stop or be rejected? Optional practice
Can we trust this training run? Optional practice
What is limiting productive compute? Optional practice
How do AI systems generate images, audio, and video? Optional practice
Who still controls a model after release? Optional practice
Choose a release mode—and plan for leakage Optional practice
Assessment & output: Work through the current scenarios and inspect feedback on your decisions. No credential is currently issued for this specialization. Core and Agent credentials are separate. See credential scope.
Before you buy
The proposed full product build and new final project are not yet available.
No date is promised for future additions. Choose based on what you can use today.
This path requires valid Core purchase history. For a first purchase, selecting a path also selects Core: US$994 total for Core + this path. You can buy Core now and add a path later.
Follow data through an AI system and practise reducing unnecessary exposure. Work through where processing happens, who may retrieve information, and what should be removed or forgotten.
What you can practise now
✓Examine compliance-readiness evidence and choices about inference location.
✓Trace data lineage, removal requirements and retrieval permissions.
✓Practise memory freshness and revocation boundaries in existing scenarios.
7 current labs · Feedback as you practise
A current exercise: What is compliance readiness?
Inspect included labs +
What is compliance readiness?
Where should inference run?
The source data is deleted. What still needs to change?
How do you design identity-bound grounding architecture?
How do you keep vector retrieval authorized, fresh, and deletable?
How should AI memory stay fresh, attributable, and revocable?
Retrieval architecture and permission test plan
Assessment & output: Work through the current scenarios and inspect feedback on your decisions. No credential is currently issued for this specialization. Core and Agent credentials are separate. See credential scope.
Before you buy
The proposed privacy attack and verification labs and complete final project are not yet available. Practice is not legal advice or a compliance certification.
No date is promised for future additions. Choose based on what you can use today.
This path requires valid Core purchase history. For a first purchase, selecting a path also selects Core: US$994 total for Core + this path. You can buy Core now and add a path later.
Explore decisions behind spoken and real-time AI interactions. The current labs cover modalities, conversation timing, handoff failures, and safe handling of tool actions.
What you can practise now
✓Compare interaction modalities and constraints in a real-time conversation.
✓Diagnose failures at handoffs between a conversation and tools.
✓Practise safe action contracts and retry decisions when outcomes are uncertain.
4 current labs · Feedback as you practise
A current exercise: What is a multimodal model?
Inspect included labs +
What is a multimodal model?
What changes when AI interaction becomes real-time?
Where did a multimodal workflow fail?
How do you prevent duplicate or uncertain agent actions?
Assessment & output: Work through the current scenarios and inspect feedback on your decisions. No credential is currently issued for this specialization. Core and Agent credentials are separate. See credential scope.
Before you buy
Full voice-application construction, the proposed new voice labs and final project are not yet available.
No date is promised for future additions. Choose based on what you can use today.
This path requires valid Core purchase history. For a first purchase, selecting a path also selects Core: US$994 total for Core + this path. You can buy Core now and add a path later.
You do not need code to spot this mistake. Here is a staff-worked example from an existing Core lab: check a short brief, repair it, then reuse the instruction when the facts change.
The task
Write a workshop brief using only the supplied notes.
Source notes
The workshop is on Thursday at 14:00 in the library. Twelve places are available. Booking closes on Tuesday.
Prepared, fictional data. Not a live AI response, customer result or account assessment.
From convincing to checked
Initial draft · one fact is wrong
The community workshop is on Friday at 14:00 in the library. There are twelve places. Book by Tuesday.
Source-backed repair
The community workshop is on Thursday at 14:00 in the library. There are twelve places. Book by Tuesday.
Friday became Thursday because the source says so—not because a different answer sounded better.
What if the source changes? +
Updated notes
Replace the previous booking details with this update: the workshop is now on Friday at 15:00. Eight places are available. Book by Wednesday. A location has not been confirmed. Do not reuse the previous room booking.
A checked brief for the new case
The community workshop is on Friday at 15:00. Location is not confirmed. There are eight places. Book by Wednesday.
The old room is not carried forward. A missing detail stays visible.
Something you can reuse
Create a short, factual workshop brief using only the supplied source notes. Include the day, start time, location, places available and booking deadline. Replace old details when the notes change. If a detail is missing, say it is not confirmed; do not invent it. Before using the brief, check every fact against the notes.
Checking these prepared cases does not establish reliability on other tasks. This is an illustration, not proof that you have completed the lab.
Support assistant / release checkEvidence incomplete
Good answers. Enough evidence?
The assistant answers the sample questions. A deletion request has been processed. Here is what the checks establish.
Answers cite the sample sources
Checked
Unauthorized access is rejected
Checked
Deleted content is unretrievable
Not checked
One deletion. Three places to check.
A synthetic document was removed from the source. Check whether its retrievable copies were removed too.
Source storeNot checked
Search indexNot checked
Answer cacheNot checked
Repair the path. Recheck the result.
First inspect the deletion path. A repair should respond to observed evidence, not a guessed cause.
Your local check record
Technical record (JSON)
Synthetic data · no assessment credit
Enable JavaScript for this local exercise. The missing check is deletion across every retrievable copy; passing answer checks alone is not release evidence.
The intended learning loopDecide → Build & test → Diagnose & repair → Defend the result
Work you can inspect. Not just a course you watched.
The target is a working capability plus the evidence behind it. Explore the anatomy of the intended deliverables.
production-evidence / outlineILLUSTRATIVE ANATOMY
The decision, with its evidence.
A system contract and release record that makes scope and remaining unknowns visible.
Scope
A bounded support assistant using approved sources. No consequential actions.
Evidence
Representative evaluations, source traceability, authorization and adverse-case checks.
Release call
Hold until required privacy and deletion checks have supporting evidence.
Limits
A passed fixture check is not proof of broad release readiness.
Ownership
The learner explains the controls, observes behavior and defends the decision.
Staff-authored contract outline, not an actual learner evaluation or completed mission.
deployment-record / outlineILLUSTRATIVE ANATOMY
What ran. What was observed.
A record ties behavior to a specific endpoint and version—not an assumption about a screenshot.
Endpoint
A learner-owned endpoint and its provider boundary would be identified here.
Version
Manifest and package identifiers identify the submitted version.
Checks
Observed behavior, verifier conditions and failed checks stay inspectable.
Limits
Observed behavior and endpoint key control do not independently attest source or complete a mission.
Teardown
A removal or shutdown record closes the operating scope.
No endpoint, hash, signed receipt or deployment result has been fabricated.
Private until you deliberately publish. Public sharing is a separate choice with privacy, secret, rights and attribution checks. Access expiry does not force publication.
You don’t need to arrive knowing how.
No coding or AI experience is needed to begin. Core offers a broad foundation; the specializations let you explore a particular interest.
Start in your browser
A laptop or desktop and an internet connection are enough to try a lab. Reading and compatible simulations also work on mobile.
You can also explore independently with documentation and an AI assistant. These labs are for people who want practice, feedback and skills they can use.
This is a learning product, not a custom development service. You do not need to buy the whole catalog to start.
Your time. Your pace. One term per product.
Each initial purchase gives you 12 months from successful payment. Products bought together run concurrently—not consecutively. Add a different path when it suits you; it does not renew any other product.
See how access fits together
Initial-purchase illustration, not a study-time estimate. Eligible early renewals extend from the existing expiry; new and post-expiry terms begin at payment.
PaymentMonth 12Month 24
Each product has its own 12-month term. Buying several together does not create consecutive years of access.
When access ends
That product’s full labs close. Account records, invoices, valid purchase history, retained artifacts, already-created exports and historical receipts remain, subject to account and privacy policies. New exports and funded verification need current access.
When you operate independently
Required completion usage is included. Keeping your system running beyond the included verification scope may create your own provider or Cloudflare costs. Renewal is optional at the then-current price; new paid paths are separate.
A few things worth knowing.
Clear scope before a commitment. Still unsure about fit, access or a purchase?
Human support Target: first response within one business day. This is not a guaranteed resolution time or contractual SLA.
What can I use today?+
Interactive laboratories, saved work and evaluation, account/evidence controls, private portable exports and observed-behavior verification are implemented for authorized accounts. Core and Agent credentials assess the named saved lab tasks; they do not claim external production deployment or universal professional expertise.
The prices and purchase interface are shown so you can inspect the actual offer. Checkout remains subject to the current payment, availability and eligibility checks. Read the current product scope or try the free experience.
Do I have to start with Core?+
Yes. Core is the foundation required for all seven specializations. Buy Core first or include it with your first path purchase. Each product costs US$497 and has its own 12-month term; products bought together run concurrently.
Already purchased Core? You do not need to buy it again or renew an expired Core term just to add a path. Valid purchase history, not an active Core learning term, satisfies the requirement. A refunded or reversed Core purchase does not qualify.
Will I need extra API credits or a paid cloud plan?+
No paid external model/API/cloud credits are required to complete the supported work. Required model, API, simulation, evaluation and managed deployment usage is included during active access. No learner credit balance or paid top-up.
The portable workflow requires a free learner-owned Cloudflare account. Optional independent operation beyond the included scope can incur your own costs.
What help and assessment retries are included?+
Learning activities provide feedback against their evidence targets. Support covers accounts, billing, access, platform problems, the official environment, included execution, rubric interpretation, accessibility, privacy and security. Platform and scoring defects can be escalated to a human.
Saved lab work is assessed directly, with no separate exam, lifetime submission quota or retake fee. Protective request limits prevent abuse without turning a service failure into a failed attempt.
Routine private artifact review, custom development, arbitrary external-code debugging and operating a post-course system are not included. Support scope and escalation.
Is there a certificate, and what would it prove?+
Core and Agent have defined credentials for demonstrated performance—not attendance or purchase. No credential is currently issued for the six other specializations. The named tasks and evidence-integrity checks must pass. Your account shows exact progress and the next task.
An issued record describes assessed versions and conditions, including permitted AI assistance. It is not an employment guarantee, checked government identity, accreditation or regulatory approval. Inspect the credential definitions.
Can my employer pay while I take the labs?+
Yes. Stripe collects the billing details; after confirmed payment you explicitly choose and confirm the learner email. The receipt email does not silently become the learner. Each assigned seat belongs to one person and is non-transferable.
Stripe handles supported tax IDs, billing details and the available payment methods. An optional PO reference can be recorded after payment; it does not create credit terms or purchase-order approval. Business billing does not include multi-seat administration or transferable seats. Support can investigate genuine billing or account-linking errors.
How do payment, refunds and location eligibility work?+
Prices are in US dollars. Stripe shows your discount and final payable total before payment. Your card issuer may charge currency-conversion fees.
Access requires successful payment and eligibility verification. An ineligible paid transaction follows the protected refund process; payment alone does not grant access. Each purchase includes a 14-day satisfaction guarantee in addition to statutory rights. Read the refund policy and purchase terms.
Does buying more products give me more years?+
No. Each selected initial product gets its own 12-month term starting at the same payment. They run concurrently. Adding another product later starts its own term and never renews the others.
An eligible early renewal extends from that product’s current expiry. New or post-expiry terms start at payment. Renewals are deliberate, optional and priced at the then-current published price. Future prices are not locked.
What happens after I select Continue to Stripe?+
Stripe handles payment and billing. If your purchase includes Core, return here to finish learner setup; a setup link is also sent to the receipt email. Buying paths for your existing Core account? They are added to that account after payment confirmation—no learner setup again.
If confirmation is pending, do not pay again. Return to your purchase confirmation or contact support. A declined payment or service error does not erase your selection. Returning from Stripe does not by itself prove that no charge occurred.
Choose what you’ll build next.
Start with the scope that fits. Your selection is preserved, and billing stays on Stripe.