An approach to learning engineering, demonstrated · about 9 minutes
Building a Learning Function
This is how I build learning functions: on intake, capacity, governance, and measurement. The proof is 21 months at Gigafactory Texas, where an instructional design team became a learning engineering function with $50M+ of documented operational impact.
- Role Learning Engineering Leader
- Scale 380,000+ training hours · 120,000+ learners
- Evaluation Kirkpatrick 1–4 + Phillips L5
- 01ProblemsEvery request was defensible; that was the problem
- 02Learning Engineering SystemsOperations before content
- 03Training and Delivery SystemsGoverned to a standard, not personally supervised
- 04LMS ArchitectureA record nobody trusts is not a delivered course
- 05CapacityCapacity is arithmetic, not advocacy
- 06LeadershipA function that cannot say no has a queue
- 07ProofClaim only what survives isolation of effects
The same problem, six environments
Multi-site education operations in Shanghai. Assessment systems across sixty-plus sections of early learners. A Texas classroom through the pandemic. Google's Privacy organization. Now the largest manufacturing building in North America. Different sectors and stakes, one repeated move: treat learning as a system with inputs, constraints, and measurable outputs, and prove the connection to real performance. That pattern is why a ramping factory read as a systems problem, not a chaos problem, and it is what this case study demonstrates.
Problems
A ramping factory generated training demand faster than any structure could absorb it.
In November 2022, I joined the training organization at Gigafactory Texas: the largest manufacturing building in North America, hiring at a pace that turned every process into a scaling problem within weeks of its being invented. Training demand arrived from every direction at once, in whatever form the requester had time to write.
Incoming training requests
Try it. Pick the request you would build first. There is no scoring and no right answer.
- and more, every week
You would build
Qualify operators on a line that has existed for a month
The line is running now. Every shift without qualified operators is a rate and a quality risk at the same time.
Every request here has a case that good. Which is exactly why the choice cannot be made on the merits of any single one.
How the gap compounded
This was an operations problem.
- With no structured intake, there was no record of what had been asked.
- With no record, there was no basis for prioritization: the next project was set by whoever asked most recently.
- With no prioritization, the questions the business actually cares about had no answer.
Seven questions a function has to answer
Try it. Select any wedge to read the question it stands for.
A ranking is only defensible if the criteria came from the business rather than from the training team.
- Priority What do we build first, and on what criteria?
- Method What should be trained at all, and what should be automated or built into the work?
- Delivery Who delivers, to what standard, and who holds it?
- Records Where does the learning live, and can the completion record be trusted?
- Roadmap What does the backlog cost us, and what will it take to clear it?
- Scope What is ours to build, and what is not?
- Impact What was the last thing we built worth?
Learning Engineering Systems
What we build first, on what criteria, and whether it should be built at all. The answer starts with the least glamorous thing in the building: a way for work to enter the team.
Intake and prioritization
I built the intake so a training request could be raised from anywhere in the factory, and every route landed in the same queue.
Multiple intake channels
A web portal, document-system forms, and request buttons inside the applications people already worked in.
Structured at submission
The form made the requester state the problem, the audience, and the business justification before a ticket existed.
Automated ticket creation
Every submission created a work item in Azure DevOps automatically, feeding a priority dashboard built for the conversation with production leadership.
The change people felt was social. Nobody walked up to a learning engineer, a training manager, or a trainer to ask for training any more. They were pointed at the intake, and the queue answered for itself.
01 Assign
Try it. Select any stage to see the rule that governs it, or switch to Before.
Web portal
A request form anyone in the plant could reach, with required fields enforced at entry, so a ticket could not exist without a requester, a business justification, and an affected population.
- Web portal
- A request form anyone could reach, with required fields enforced at entry.
- SharePoint forms
- The same intake, embedded where several departments already kept their documentation.
- In-tool buttons
- Request actions built into the internal applications people already used during the work day.
- Ticket, automatic
- Every submission created a work item in Azure DevOps with no manual re-entry.
- Priority dashboard
- Backlog, live queue, ranked priority, and roadmap in one view, synced with the ticket system.
- Weekly digest
- An automated summary of active projects and timelines to stakeholders.
- 01 Assign
- Each released ticket goes to a named learning engineer who project-manages it end to end. I managed the portfolio rather than the projects.
Deciding what to build
Every released ticket went to a named learning engineer who project-managed it end to end. I managed the portfolio and held the quality gate. What those engineers did at stage 02 was not design a course. It was performance analysis, run on the Six Boxes® model, which treats training as one of six influences on performance rather than as the answer. Five of the six sit outside anything a course can touch.
02 Analysis
Try it. Select any box to see what a gap there looks like and what closing it takes.
Expectations and feedback
The gap: people are not certain what a good output looks like, or they learn that it was wrong long after they could have acted on it. Closing it: published standards, visible targets, and feedback close enough to the work to change the next one. Usually the cheapest box to fix and the most commonly skipped.
- Expectations and feedback
- Published standards, visible targets, and feedback close enough to the work to act on.
- Tools and resources
- Job aids, better instrumentation, forcing functions that make the error impossible, or automating the step out entirely.
- Consequences and incentives
- Aligning what gets recognized, measured, and escalated with the behavior the business wants.
- Skills and knowledge
- The only box a course can reach: instructional design, practice to fluency, assessment against real performance.
- Selection and assignment
- Hiring criteria, role definition, and which role the task belongs to.
- Motives and preferences
- What the work costs the performer, and whether it fits what they value.
Design and delivery method
ADDIE, with a gate
Stages 01 to 07Try it. Select any stage to see what it produces. Assignment and analysis are covered by the figures above.
03 Design
Objectives written against the performance gap, assessment designed before content, and the structure of the pathway decided. The evaluation instrument is selected here, which is why measurement is never retrofitted later.
- 03 Design
- Objectives against the performance gap, assessment before content, and selection of the evaluation instrument.
- 04 Develop
- Production against the signed design, with AI-assisted drafting held to that design.
- 05 Review gate
- SME sign-off on technical accuracy, portfolio sign-off on instructional quality.
- 06 Implement
- Handoff to instructors and training coordinators with facilitation guides and standards.
- 07 Evaluate
- Kirkpatrick 1 and 2 as standard, deeper evaluation on selected programs.
Training and Delivery Systems
Who delivers, to what standard, and who holds that standard. The team built; the training organization delivered; I owned the difference.
Governing delivery without performing it
The team creates curriculum and online courses; instructors and training coordinators deliver them, to a standard I defined and inspected. That standard was concrete: role definitions, standard operating procedures, service levels, leveling guides, facilitation guides, and train-the-trainer enablement across 40+ instructors. Scheduling and rostering stayed with the training organization, which is what kept the governance credible rather than territorial.
The team's charter stated what it does not do as explicitly as what it does. Publishing exclusions is unpopular and load-bearing: a function that will not say what it declines cannot defend what it prioritizes.
Structuring the training organization
Delivery capability does not appear on its own. I defined the trainer roles, wrote the standard operating procedures and service levels the delivery teams ran against, built the leveling guides that made progression legible, trained the trainers, and managed the programs they delivered.
The output was not a set of courses. It was an organization that could run them consistently without me in the room, which is the only version of scale that survives a leader changing jobs.
LMS Architecture
Where the learning lives, and whether the completion record can be trusted by the people who depend on it.
Systems as governed infrastructure
LMS and LXP governance across Intellum, Canvas, and StudioEdX: how courses are structured, versioned, and retired; who is permitted to publish; and how completion records connect to the qualification requirements the floor actually runs on.
A course nobody can find, or a completion record nobody trusts, is not a delivered course. Systems governance is where most learning organizations quietly lose the credibility their content earned.
Capacity
What the backlog costs, and what it takes to clear it. Not a matter of conviction: a matter of arithmetic.
Multiplying output before adding people
Before this model was used to argue for people, it was used to argue against needing them. I architected AI-assisted production behind human review gates, so engineers stopped producing first drafts and started editing them.
Draft generation
Structured course designs rendered into first-draft deliverables.
AI voice generation
Narration produced without studio time or re-record cycles.
AI video and visuals
Presenter segments and visual assets without a shoot or a queue.
Automated evaluation
Assessment items drafted from the objectives already written at scoping.
The gate never moved. Every generated asset cleared the same SME and portfolio review as anything built by hand, because a faster wrong course is still a wrong course. That is also why the leverage is bounded: it reaches drafting, not judgment. Doubling drafting speed does not double a function's capacity, and a model claiming otherwise does not survive its first quarter.
Modeling capacity
So I built the function's capacity and workforce model. It measures in effort hours, because a course is not a unit of work: a short online module and a multi-day technical curriculum differ by an order of magnitude, and a model that counts them as one thing each will lie to you at precisely the moment the answer matters.
Capacity
Headcount × annual productive hours: gross hours net of PTO, holidays, and the interruption load every real job carries.
Demand
Three terms, not one: development backlog, recurring intake, and maintenance on the existing catalog.
Net progress
Capacity minus intake minus maintenance. Clearance time is backlog divided by net progress.
The capacity model
Illustrative modelTry it. Raise production leverage first, the way the function did. Then move headcount, and watch how much of the gap the cheap lever had already closed.
- Effective capacity
- 16,224 h / yr
- Net progress
- 7,134 h / yr
- Clearance
- 1.5 years
- Maintenance
- Intake
- Available for backlog
With JavaScript enabled, six sliders drive this model live. At the default values shown, effective capacity is 16,224 hours per year, maintenance and intake consume roughly 56 percent of it, and net progress clears an 11,000-hour backlog in about 1.5 years. Raising production leverage lifts the draftable half of build work only, so a doubling of drafting speed yields about 1.5 times the capacity, not twice. Reduce headcount or raise the maintenance share far enough and net progress turns negative, at which point clearance is undefined.
Maintenance is the term that changes the conversation. It is absent from most training plans, and it scales with the catalog rather than the team: the more a function has built, the more of its capacity is consumed before any new work begins. A mature catalog can carry a sustainment load approaching a full team's worth of hours, none of it on a roadmap.
Below a threshold, the backlog is not slow to clear. It is structurally unclearable.
No prioritization scheme crosses that line, and neither does working harder. A model that locates the threshold converts a staffing conversation from advocacy into arithmetic. Capacity became an argument instead of a feeling.
Leadership
What is ours to build, and what is not. The boundary, and the decisions that made it hold.
Scope: which programs, and which responsibilities
Not all inbound demand belonged to this function. Scope was set against measured production outcomes: technical and shop programs, onboarding, leadership, and the external pipelines supplying production associates and technicians from community colleges and partner institutions. Everything else remained with its owning function.
Scope of work
Scope of authority
The other axis responsibility, not programsThree management decisions
01 Where the function sits
The team is embedded in Production, reporting through Training Operations, rather than in HR. That placement is why training priorities are set against production consequence, and why the measurement conversation happens in operational units. Proximity to the work is a governance choice, not an org-chart accident.
02 A shared language for who owns what
Where several groups deliver training into the same population, the recurring conflict is rarely about quality. It is about jurisdiction, and it surfaces as duplicated effort, as gaps everyone assumes someone else is covering, and as programs measured against outcomes they were never designed to move.
I authored a typology that defined, for each category of training, four things: its purpose, its measurement unit, its accountable owner, and the frame in which its return is claimed. Making the measurement unit part of the definition is what did the work. Once a category carries its own evidence standard, "who owns this" becomes answerable by asking what the training is supposed to change, and the conversation moves from territory to design.
03 Two management layers
Up to 12 people across Learning Design Engineers, Media Engineers, and Content Producers, including senior individual contributors carrying direct reports. Building the second layer was the step that converted a team I ran into a function that runs without me in every room.
Image placeholder 02 · optional animation
Governance motif: an abstract two-layer structure resolving from scattered nodes into a bounded system. Recreated from scratch in the token palette. A short looping SVG or CSS animation works here; it must be decorative only and pause under reduced-motion.
Governing under pressure
A queue with rules tells some requesters "not yet," and no amount of good design makes that popular on first contact. What made it hold: the triage criteria were the business's own, the rules were published rather than applied case by case, and the capacity model made every deferral a visible tradeoff instead of a judgment.
People argue with a decision. It is much harder to argue with arithmetic they can inspect.
Proof
What the last thing we built was worth, and how the number survives scrutiny. A function that cannot answer this is funded as an expense forever.
How the $50M+ was documented
Instrumentation is tiered, because not every course earns the same instrument: reaction and learning as a standard across everything, behavior and results on selected programs, and full ROI only where isolation of effects can actually be performed. That architecture is what produced the figure, documented using Phillips and Kirkpatrick methodology across four benefit streams.
- Vendor displacement Work brought in-house against market rate.
- Time to competency Reduction in the interval from hire to autonomous work.
- Failure prevention Safety-critical failures avoided.
- Production gains Non-conformance, rework, and takt or cycle time.
Where training was one contributor among several, attribution was discounted before the benefit was claimed, not after it was challenged.
A conservative number that survives scrutiny is worth more to a function than a large one that does not.
What generalizes
Learning becomes real when it is built from intake, capacity, governance, and measurement: this is what that looks like at industrial scale.
Riley Inks
Learning Engineering Leader
Every visual was drawn from scratch. Figures in the capacity model are illustrative. Published figures are limited to those already public.
← All work · rileyinks.com · Writing · Resume