Become an AI Expertv2.1
0 of 0 steps Continue
Stage 4 · Getting inFrom week 13

Module 12: The internal move and public work

Prerequisites and measurable objectives

Prerequisite: Milestone C. By the end you can:

  1. Pitch and deliver a 30-day AI pilot at your current job with a dollar cost, eval numbers and failure cases.
  2. Publish a public record (three posts, one runnable repo, one talk) that a stranger could refer you on.
  3. Write a README that a reviewer of 2026 portfolios would call finished.

Primary resource: Janvi Kalra on The Pragmatic Engineer, From Software Engineer to AI Engineer (https://newsletter.pragmaticengineer.com/p/from-software-engineer-to-ai-engineer). Finish it before adding anything from the reading list. Hours: 10 guided, which the concept blocks above account for in full; the lab run and the build fit inside them.

Why this is the highest-probability route

Of the twenty-plus people in the vetted stories, nine got into ML by taking an adjacent job and moving sideways inside the company, including the DeepMind and FAANG staff hires. A senior MLE in August 2026 says ML roles are not junior-friendly and this is how most people become MLEs. Janvi Kalra, now at OpenAI, was rejected by her company's AI team, became its go-to LLM person through side projects, and was an easy hire when it expanded: "it's much easier to transfer internally." Every verified referral came from someone who already knew the person's public work.

Concepts, in order

  1. Find the unsolved problem (1 hour): something your company pays for in human hours that your Stage 3 skills can do: triage, extraction, search over internal docs, summarising tickets. One person automated a boring part of his job and was invited onto the ML team; another pitched a model on the company's energy data as a money-maker.
  2. Pitch a pilot, not a career change (1 hour): one named problem with a dollar cost, one tool, 30 days, a budget under a few hundred dollars, the data-security question answered first. Position as a senior engineer adding AI, not a junior ML hire. A backend-shaped first project is an internal LLM gateway with timeouts, retries, idempotency, cost tracking and tracing; companies now staff such teams.
  3. Deliver with evals from day one (2 hours): the Module 08 harness on the pilot; report results with eval numbers, cost, and failure cases; expect role blur and be ready to throw work away when a new model makes it obsolete.
  4. Ask for the title (1 hour): after one successful project, ask for AI in your title or a transfer; if the company has nothing AI-adjacent, move to one that does in your current role and repeat.
  5. Public work (3 hours): a free GitHub Pages blog; three posts with numbers (the nanochat write-up, the cost cut, the reversed decision); one runnable repo; one talk at a meetup or recorded; one upstream contribution; answering questions in a community weekly. The DeepMind hire built a YouTube channel and open-source repos per topic for 1.5 years. Generosity created the opportunities in the most upvoted self-taught story.
  6. The README that reads as finished (1 hour): one-line result with a number; architecture diagram; eval table against a baseline on a described, versioned set; p50/p95 and cost per request; a "where this breaks" section; reversed decisions; three commands to run. Make clear what is yours; never post course exercises as projects. Start from templates/project-readme.md.
  7. Contributing upstream without wasting a maintainer's time (1 hour): "good first issue" labels; doc drift as a first PR; RFC before large changes; disclose AI use; review every line yourself. Maintainers spot AI-generated PRs.

Lab

No lab. The artifact is the pilot and the public record.

Build

Project 12: the pilot. A pitch document (the pilot-pitch template in templates/pilot-pitch.md) with the dollar cost; the delivered pilot with its harness; a results memo with eval numbers, cost and failure cases; the title conversation had.

Project 13: the public record. Three posts, one runnable repo with the finished README, one talk, one upstream PR opened.

Deliverables: the pitch and results memo (redacted if needed); links to the posts, repo, talk and PR.

Mastery gate

A person outside your company has starred, cited or commented on your work, and your pilot results memo has eval numbers your reviewer can trace.

Quiz

  1. What are the three parts of a pilot pitch that get approved, and what gets one rejected?
  2. Why position as a senior engineer adding AI rather than a junior ML hire?
  3. What belongs in the first line of a finished README?
  4. Why does doc drift make a good first upstream PR?
  5. Name three things reviewers of 2026 portfolios say make a project read as unfinished.
  6. What did the DeepMind hire do in each three-month block, and why did it work?

Primary-source reading

Follow-ups

When you finish this module, paste one of these into your assistant. In Claude Code write /aiexpert-tutor before it; in Codex write $aiexpert-tutor.

  1. Review my pilot pitch as the sceptical manager who has to fund it.
  2. Ask me what a stranger would cite in my public post and whether every number in it is traceable.
  3. Rehearse the title conversation with me. You play the manager.

Free to use: code MIT, writing CC BY 4.0, third-party media under its owners’ terms. Licence and terms.