The Residency Model

Safua is structured as a three-phase engineering residency: Foundation, Build, Prove. The shape is borrowed deliberately from medical training. That choice is load-bearing — it's not decoration.

This page explains why.

Why a residency, not a bootcamp or a degree#

Engineering education has three dominant shapes, and each of them is breaking for AI engineers specifically.

Four-year degrees were designed for a world where the curriculum had a longer half-life than the degree itself. That condition no longer holds in AI. A model architecture taught as state-of-the-art in a freshman course is a historical artifact by the time the student graduates. The structural pacing of a degree — semester-long courses, summative exams, GPA as the signal — is optimised for knowledge retention. AI hiring is not asking for retention. It's asking for current engineering taste against tools that shipped three months ago.

Bootcamps compressed the tempo but kept the shape: cohort-locked, instructor-led, syllabus-driven, exam-graded. The best ones were excellent. Most collapsed. The economics — short programmes, narrow margins, dependency on placement outcomes for revenue — made them fragile when the job market tightened. MOOCs went the other way: async, unlimited enrollment, no placement promise. The tradeoff is completion rate: 12.6% across the category, and that's the optimistic number.

Residency is a different shape entirely. It's the model medicine uses for the jump from classroom-certified to trusted-with-a-patient.

The structural commitments of a residency:

Supervision, not instruction. Residents aren't taught by lecturers. They do real work, supervised by senior practitioners who critique the work as it happens.
Variable duration. You finish when you're ready, not when the semester ends. Progression is skills-gated, not time-gated.
Public record. Every rotation, every case, every documented incident is part of the resident's permanent record. When they graduate, that record is what employers evaluate.
Named accountability. The supervising engineer owns the feedback. There's no "the department" or "the grading rubric" as the authority — there's Dr. Mendes or Dr. Kowalski, and their review is signed.

Applied to engineering, that's Safua.

Foundation: calibrate, don't teach#

Foundation is the first phase of the residency. Its purpose is narrow and non-negotiable: figure out what the learner actually knows, and route them into Build at the right level.

It is not a course. There are no lectures. There is no "Python 101." What Foundation does is probe — a small number of carefully-scoped tasks across six domains (Python, Git, APIs, SQL, Docker, AI integration) — and use the responses to place the learner.

The reasoning behind this shape: most learners come to Safua with uneven skills. Strong SQL, weak Docker. Fluent in classical ML, shaky on production serving. If the platform treated them as a blank baseline, Build missions would be miscalibrated — too easy in their strong domains (no growth) and impossibly hard in their weak ones (frustration, drop-off).

The output of Foundation is a learner-specific skill vector, not a grade. That vector feeds into mission selection in Build.

Build: ship real work under review#

Build is the long middle of the residency. It's where almost all platform time is spent.

The unit of work is a mission. A mission is a scoped engineering brief — the kind of task a senior engineer would hand a junior if the team were short-staffed and the senior engineer trusted the junior to ship something real. Missions are anchored in one of four virtual companies so the review priorities reflect real industry standards, not generic "code smell" heuristics.

Submission, review, iteration. Every mission goes through that cycle. A mission isn't complete when the code compiles. It's complete when the named reviewer signs off on the iteration.

The reviewer is a specific engineer on the faculty. Their feedback is written in their voice, scored against Safua's five-dimension rubric, and added to the learner's permanent record. The reviewer doesn't change between submissions — a learner who's shipping Data Engineering missions at NovaMind builds a review relationship with Rafael Mendes, and that continuity is part of the pedagogy.

Build is the phase that looks most like a job and least like school. That's the whole point.

Prove: graduate into a public record#

Prove is the shortest phase in calendar time but the most consequential in outcome. It's the step where a learner's Build history becomes a public, verifiable profile at safua.ai/u/<handle>.

To graduate into Prove, a learner has to clear a school-specific Confidence Score threshold. The threshold varies by school — AI Safety & Governance sets a higher bar than, say, Data Engineering, because the downside of a weak safety engineer is worse than the downside of a weak ETL engineer.

Once graduated, the learner's profile aggregates:

Every mission: brief, reviewed submission, iteration history
Every score: per-dimension, per-mission, and aggregate
Top artifacts: reviewer-flagged highlights
Emerging strengths: the pattern the reviewers themselves surfaced

The profile is the credential. Not a PDF. Not a badge. The work, the review, the reviewer's name, the date.

Because every review is attributed to a specific named engineer on the Faculty page and the submission log is append-only, a profile entry cannot be added or altered without an identifiable faculty signature behind it. That property is why third parties can verify the record directly — they don't have to trust a claim; they can read the ledger.

What this shape gives up#

The residency model is not strictly better than the alternatives. It has real tradeoffs. Being honest about them:

No fixed duration. Some learners finish in six months. Some take eighteen. If a learner needs a "I'll be done by May" commitment, Safua is the wrong fit — the pacing is skills-gated.
Higher cognitive load. Supervised iteration with named feedback is more demanding than watching lectures. Learners who want passive consumption leave within two weeks. That's working as intended.
Reviewer cost. Named faculty reviews are more expensive to run than automated grading. Safua's pricing reflects that; the pricing page is explicit.

We don't apologise for any of these. They're the structural features that make the credential worth something.