Specialization Tracks¶
Ninety-day, project-shaped tutorials that reuse the core modules. These are not phase checklists alone — each track page has mental models, explainers, code sketches, traps, and exit gates.
Complete Setup first. Close core gaps listed on the track before you start building.
Tracks are domain-specific; the Capstone is domain-agnostic and can run instead of or alongside a track — it proves the six generic parts (service, evals, knowledge, agent, ops, security) hold together, without committing to a vertical.
| Track | Outcome | Core dependencies | Vibe |
|---|---|---|---|
| Stock recommender | Research assistant / recommender prototype: data → baseline ML → SLM → RAG → compression → ship | 01–07, 09–10, 13–14, 17, 23; 22/24 if you add a tool loop | Markets + retrieval + MLOps (not financial advice) |
| Hybrid models | Custom Transformer + MLP fusion in PyTorch, ablations, deploy | 05–06 + DL fundamentals; 23 analog (config pins), 04/22 regression helper | From-scratch architecture engineering |
| Agentic editor plugin | VS Code extension + agent backend + local models | 01–05, 07–08, 11–12, 17, 20–25 | IDE product + agent safety |
How to run a track (intuition-first)¶
flowchart LR
Setup[Setup + core gaps] --> Story[Read track story + architecture]
Story --> Phase[One phase at a time]
Phase --> Lab[Ship phase exit artifact]
Lab --> Log[PROGRESS.md + honest metrics]
Log --> Phase
Log --> Demo[Day-90 demo + rubric]Day 1: start from the slice in tracks/starters/, not an empty repo and not a finished 90-day demo.
- Skim the whole track once — know the day-90 shape before day 1.
- Read the track’s Intuition lock out loud; if you can’t restate it, you’re not ready to code.
- Map each phase → core modules; finish those modules’ labs first.
- Ship phase exits (repo artifacts), not only notes.
- Score yourself with Assessment rubrics at mid-track and day 90.
- Optional: keep Progress XP for core modules; tracks are graded by demos.
Shared engineering bar¶
| Bar | Why |
|---|---|
| Git from day 1 | Tracks die in “works on my laptop” folders |
| Tests for deterministic code | Splits, tools, parsers, shapes |
| Evals for model behavior | Golden Q&A, must-refuse, Hit@k — not vibes |
| Agent hardening (if you ship a loop) | Failure detectors, sandbox/HITL, trajectory eval, prompt pins (modules 20–23) |
| Hardware honesty (local models) | Size to RAM/KV (17 §7); do not swap |
| README with limits + ethics | Especially finance and write-capable agents |
| No secrets in git | Keys in env / SecretStorage only |
| Time / data leakage honesty | Shuffle is cheating on markets and sequences |
Choosing a track¶
| If you want… | Pick |
|---|---|
| End-to-end product with data + RAG + deploy | Stock recommender |
| Deep understanding of architectures and training | Hybrid models |
| Shipping developer tooling with agent safety | Agentic editor plugin |
You may run a track in parallel with later core modules if you already ship Python/TS services confidently.
Which new core patterns belong where¶
Modules 20–27 (the agent-hardening tail of Gate 4 and Gate 5) and 17 §7 limited hardware are not a tax on every track. Cargo-culting LangGraph, worktrees, or trajectory evals onto a tabular hybrid is how you get costume jewelry.
| Pattern | Stock recommender | Hybrid models | Agentic plugin |
|---|---|---|---|
| 17 §7 RAM / one resident model | Yes — PEFT + lite serve | Partial — shrink d_model / max_len / batch, not GGUF |
Yes — Ollama default |
| 23 Prompt/config digest | Yes — research prompt pack | Yes analog — train YAML + metrics JSON | Yes — system + tool list |
| 22 Trajectory / regression | Optional (only if you add a tool loop) | eval_regression on MAE, not agent traces | Yes — stubbed agent CI |
| 24 Token budget / local-first | If /research calls an SLM in a loop |
No | Yes |
| 20–21 Breakers, manifests, worktrees | Quote-tool timeouts; not a coding agent | No | Yes — this is the incident |
| 25 Durable HITL / merge gate | No | No | Yes — approve then apply |
| 26 Orchestrator comparison | Optional cost-per-path | Params/latency log, not CrewAI | Written custom vs LangGraph |
Day-90 definition of done (all tracks)¶
- Demo runs from a clean clone + documented setup
- At least one automated test suite and one model/behavior eval
- Architecture diagram in README matches the code
- Known failure modes written down (not hidden)
- Ethics / safety / non-advice notes where relevant
- Patterns from the table above that apply to this track are in the demo, not only in notes
Next: open a track page and start with its story + system diagram.