Curriculum Audit-tutoriaal¶
This document is the committed analysis behind the DEEP_DIVES enhancement. Re-read at every quarterly retrospective; update when reality contradicts an assumption.
What tutoriaal is¶
A 12-month, ~12-hour-per-week applied-AI curriculum with three layers:
- `AI_EXPERT_ROADMAP.md - strategic doc establishing identity (AI Engineer with eval/agent/observability specialty + AI infra moat), KPIs, anti-patterns.
- `sequences/ - 17 topical files. Each has rungs (sub-skills) with "done when" gates.
- `weeks/ - 48 week files. Three-session cadence (Theory / Build / Synthesis).
- `DEEP_DIVES/ - 14 self-contained reference chapters added during the hardening pass.
Strengths to preserve¶
- Identity-first framing.
- Evals-first discipline.
- Public-default cadence (repo on day 1, blog post per month, OSS PR per quarter).
- Three-session-per-week structure (deep blocks beat fragmented dailies).
- Honest cadence handling (24-month variant if 6 hr/wk).
- Anti-pattern list (diagram theatre, mock-it-out, tool tourism, scope laundering).
- Anchor projects per quarter.
- SRE-as-bridge-not-cage framing.
Audit findings (the gaps the DEEP_DIVES patch)¶
| Gap | Patched by |
|---|---|
| External-link rot risk | All 14 chapters are self-contained; no chapter requires YouTube/Karpathy/Strang to learn |
| Math taught only via 3B1B link | DEEP_DIVES/01 |
| PyTorch user-level depth | DEEP_DIVES/02 |
| Classical ML rigor often skipped | DEEP_DIVES/03 |
| Backprop / optimizer derivation | DEEP_DIVES/04 |
| LLM application patterns at survey only | DEEP_DIVES/05 |
| RAG without metric derivations | DEEP_DIVES/06 |
| Agents without distributed-systems lens | DEEP_DIVES/07 |
| Eval methodology (the user's specialty) shallow | DEEP_DIVES/08 |
| LLM observability (the user's moat) shallow | DEEP_DIVES/09 |
| LoRA/QLoRA/DPO papers referenced but not derived | DEEP_DIVES/10 |
| Multimodal absent (text-only curriculum) | DEEP_DIVES/11 |
| Prompt-injection / red-teaming as one bullet | DEEP_DIVES/12 |
| AI-for-SRE direction underweighted | DEEP_DIVES/13 |
| No durability/refresh discipline | DEEP_DIVES/14 |
Future-proofing verdict¶
Spine (math, transformer fundamentals, evals discipline, distributed-systems thinking applied to agents): 10+ year half-life. Durable.
Stable (specific architectures, well-published algorithms): 4-7 year half-life. Refresh annually.
Ephemeral (tool versions, framework APIs, vendor features, model names, pricing): 1-3 year half-life. Refresh quarterly.
The original tutoriaal mixed all three without distinction. With DEEP_DIVES + chapter 14's audit framework, the reader can refresh appropriately by tier.
Market realism (2026)¶
- The eval/agents/observability lane is real and undersupplied.
- 12 months of disciplined work + prior backend experience produces a genuine applied-AI engineer.
- Salary band for the curriculum's claimed top-tier ($300-700k) is the 75-90th percentile, not median. The median for the bridge profile is closer to $200-400k in 2026 US markets.
- The capstone trio (eval framework + observability post + capstone repo) is interview-credible.
Future-market hedge (2027-2030)¶
The curriculum's structural premise (1-year-to-credible) is sound; the specific track choice must be re-evaluated yearly per chapter 14's pivot signals. Future-proofing depends on:
- The DEEP_DIVES being the durable layer (math, derivations, design patterns).
- The sequences/weeks being the time-grained layer (refresh-as-you-go).
- The audit (this document) being re-read annually.
Decision rules going forward¶
- Yearly: re-read this AUDIT.md and DEEP_DIVES/14. Update findings. Decide: continue, deepen, or pivot.
- Quarterly: refresh one quarter's sequences. Verify DEEP_DIVES still match reality.
- On significant field shift: rewrite affected DEEP_DIVE chapters; commit dated updates.
Acceptance criteria for "the enhancement worked"¶
- By end of year 1: at least one DEEP_DIVE chapter has been updated based on personal experience.
- By end of year 1: external-link rot has been mitigated (existing sequences updated to reference DEEP_DIVES first, links second).
- By end of year 1: chapter 14 has been re-read at least once.
- At quarterly retros: the relevant DEEP_DIVE chapter is the primary reference, not a YouTube link.
If any of the above is false at year-end retrospective, the enhancement underdelivered and the system needs further work.