Back to Home
Living document

Road to AI Engineer

Full-stack today, AI engineer by 2029. Every phase, every project, every habit — and the parts I get wrong. Updated as I go, not written after the fact.

Read the full research postLast updated: September 5, 2026
Current phase
Phase 0Current phase
Portfolio projects shipped
0 / 7Portfolio projects shipped
Skill gaps tracked
10Skill gaps tracked
Target senior profile
2029Target senior profile

What I'm doing right now

The honest snapshot. If it isn't on this list this week, it isn't happening this week.

  • Running plan → worktrees → verification → risk-based review on a real client project, not a toy repo.
  • Writing and maintaining CLAUDE.md, hooks and permission allow-lists for every repo I touch.
  • Working through Anthropic Academy: Claude Code in Action, MCP intro and advanced, Agent Skills and Subagents.
  • Measuring time-in-review for four weeks so Phase 0 ends with a number, not a feeling.
  • Reading Anthropic's engineering posts on effective agents, context engineering, evals and harness design.

The phases

Sep 2026 → 2029. Each phase has an end condition, not just an end date.

  1. Phase 0Sep – Oct 2026In progress

    Work like the elite at my current job

    Zero cost, immediate return, and the first line of an AI engineer résumé. Change how I work before changing what I study.

    35%
    • Adopt plan → worktrees → verification → risk-based review on a real project
    • Write and maintain the repo's CLAUDE.md
    • Set up hooks, permission allow-lists and a /commit-push-pr command
    • Add AI review before human review, measure time-in-review for four weeks
    • Anthropic Academy: Claude Code in Action, MCP, Agent Skills and Subagents
    Claude Code
    Agent harness
    CI review
  2. Phase 1Nov 2026 – Feb 2027Planned

    Fundamentals in my stack + first real project

    Learn the API layer properly in TypeScript, add production Python, then ship one retrieval system that a real business actually uses.

    0%
    • Anthropic Academy — Building with the Claude API (8h)
    • Matt Pocock's AI SDK v6 Crash Course: agents, tool calling, eval-driven development
    • Chip Huyen — AI Engineering
    • Production Python: FastAPI, Pydantic, async
    • MCP 2026-07-28 spec, end to end
    • Project 1: RAG with hybrid retrieval, reranking, citations, 50-question eval set, Langfuse from day one
    Claude API
    AI SDK v6
    FastAPI
    RAG
  3. Phase 2Mar – Jun 2027Planned

    Evals and agents — the differentiator

    The part most portfolios skip, and the part senior interviews are actually built around.

    0%
    • Hamel Husain's free evals material + the evals-skills repo
    • W&B and DeepLearning.AI eval courses, Andrew Ng's Agentic AI
    • Claude Certified Developer exam
    • Project 2: eval harness with a labeled failure taxonomy and a CI regression gate
    • Project 3: MCP server for a real system, OAuth 2.1
    • Project 4: back-office agent on the Claude Agent SDK, approval hooks, documented cost per run
    • AI Engineer Miami — April 26–27, 2027
    Evals
    LLM-as-judge
    Agent SDK
    MCP servers
  4. Phase 3Jul – Dec 2027Planned

    Depth, visibility, first applications

    Stop learning in private. Contribute, publish numbers, and start applying to remote AI product engineer roles.

    0%
    • Hugging Face Agents and Context courses, LangGraph intro, Google's agents intensive
    • Project 5: document extraction pipeline with a human review UI
    • Project 6: voice agent for a hackathon
    • Project 7: internal copilot with cost and latency dashboards
    • Contribute to MCP, Langfuse, Promptfoo or Mastra
    • Publish three technical posts with real numbers
    • Start applying to remote AI product engineer roles
    Open source
    Writing
    Job search
  5. Years 2–32028 – 2029Planned

    Specialty and senior profile

    Understand models from the inside, pick a specialty, and talk about a system of mine that runs in production.

    0%
    • Stanford CS336 and Karpathy's nanochat
    • Raschka's two 'from scratch' books
    • Berkeley Agentic AI MOOC
    • Claude Certified Architect track, if I move toward architecture or consulting
    • A conference talk about a production system of mine
    Fundamentals
    Architecture
    Speaking

My gap, honestly

Ordered by how often hiring sources cite it. The bar is where I am today, not where I want to be.

  • Evals and error analysis

    15%

    The single most cited gap. Golden datasets, LLM-as-judge validated against human labels.

  • Retrieval quality

    30%

    Not just 'RAG': hybrid search, reranking, naming failure modes unprompted.

  • Context engineering

    40%

    Intentional compaction and the long-context dumb zone.

  • Bounded agent design

    35%

    Including knowing when a plain workflow beats an agent.

  • Cost and latency economics

    30%

    Model routing, caching, structured outputs.

  • LLM security

    20%

    Prompt injection, exfiltration through tools, OWASP LLM Top 10.

  • Observability

    25%

    Tracing, prompt versioning, drift, cost monitoring.

  • Production Python

    35%

    FastAPI, Pydantic, async. My stack is TypeScript-first.

  • Document pipelines

    20%

    Chunking, OCR, extraction, deduplication.

  • Fine-tuning basics

    10%

    Low priority: mostly knowing when to pick it over RAG or prompting.

The seven projects

Each one ships with an eval suite and traces — that combination is the portfolio.

  1. Planned

    RAG over a real business's documents

    Hybrid retrieval, reranking and citations over documents a business depends on — not a demo corpus.

    Next.js
    Vercel AI SDK
    pgvector
    Promptfoo
  2. Planned

    Eval harness for that RAG

    An LLM judge validated against human labels, with a labeled failure taxonomy.

    Promptfoo
    Langfuse
    CI
  3. Planned

    MCP server for a real system

    OAuth 2.1, published to a registry, used by something other than my own laptop.

    MCP
    OAuth 2.1
    TypeScript
  4. Planned

    Back-office agent

    Invoices → extraction → validation → ERP, with approval hooks and documented cost per run.

    Claude Agent SDK
    Hooks
    ERP
  5. Planned

    Document extraction pipeline

    Confidence scores and a human review UI, because extraction without review is a liability.

    Python
    FastAPI
    OCR
  6. Planned

    Voice agent with tool calling

    Built for a hackathon, so the deadline is external.

    Realtime API
    Tools
  7. Planned

    Internal copilot

    OpenTelemetry GenAI spans, cost and latency SLOs on a dashboard someone else reads.

    OTel
    Grafana
    Agent SDK

Day to day

The habits underneath the roadmap. Boring on purpose.

  • Ship in the open

    Every project gets a repo, a README with real numbers, and a post. Private work teaches me; public work gets me hired.

  • Evals before features

    No new capability lands without something that can tell me it regressed. Vibes are not a test suite.

  • One real user

    Every project points at a real business or a real person. Demo corpora hide the failure modes that interviews ask about.

  • Measure the boring number

    Time-in-review, cost per run, p95 latency. Claims without numbers are opinions.

  • Agents on real repos

    Plan, worktrees, verification, risk-based review — on production work, daily, not on toy problems.

  • Write the post while it's fresh

    Notes decay fast. This page exists so I can't quietly rewrite history later.

Following / ignoring

What I actually act on, and the noise I decided to skip. Both cost time.

Following

  • Evals as the hiring filter

    Moving beyond vibes: golden datasets, LLM-as-judge validated against humans, regression suites gated in CI.

  • 'What has gone wrong for you?'

    The question that opens almost every senior interview. Worth having three real answers ready.

  • Time zone as the real advantage

    UTC-6 is US Central. Every nearshore guide lists it first, and Honduras appears in none of the LATAM AI-talent guides.

  • Buy-over-build enterprise reality

    76% of enterprise AI use cases are bought, not built. Someone has to integrate them into real processes — that's the job.

Ignoring

  • 'A million lines in a week'

    Teams with verifiable throughput credit linters, docs, tests and sandboxes — not fleets of 30 autonomous agents.

  • Self-reported adoption stats

    '98% of employees use X' and '47% of code is agent-written' are perception or vendor data.

  • $450K FDE salary screenshots

    Frontier labs only, 55–70% in equity, and mostly requiring US/UK/EU presence.

  • The $4,200 evals cohort

    Out of budget, and the free material covers most of it. Cloud certs only if an employer's stack demands them.

Progress log

Short entries as things land. Newest first.

  1. Roadmap goes public

    Published the research post and turned it into this page. Making the plan visible is the cheapest form of accountability I know.

    Phase 0
    Writing

  2. Phase 0 starts

    CLAUDE.md written for the main client repo. Hooks and permission allow-lists next, then four weeks of time-in-review measurements.

    Phase 0
    Claude Code
Road to AI Engineer — Angel Valladares