Software engineer to AI engineer: a build-first roadmap

The table covers five postings we read on October 23, 2026: a sample, not a survey. Pay is each posting's listed range, with its own label where it gives one.

Table 1. Scroll sideways for every column.

Three patterns hold in this sample. Every posting names evaluation, and four name retrieval or RAG. None asks for a PhD, and BLEN says its role is not research and does not involve training foundation models ²; Twin Health's staff role, though, lists building deep learning models among its duties ⁵. The experience bar varies most, from early-career with projects to 8+ years. Read the requirements, not the title.

Much of what the postings ask for is ordinary software engineering: production Python backends, APIs, SQL, a cloud platform, Git, testing and CI/CD ^2,3,4,5. Four things are likely to be new. The postings repeat evaluation, retrieval and model APIs; designing for variance is the habit that ties them together.

• Evaluation. OpenAI says generative AI is variable, so the same input can produce different output and traditional software testing is insufficient ⁷. A 2024 study of five models configured to be deterministic found accuracy varied by up to 15% across naturally occurring runs ⁸. Anthropic advises specific, measurable success criteria and evals that mirror real use, include edge cases and favor many automatically graded cases over a few hand-graded ones ¹⁰. If a model grades the output, validate it against human labels, as OpenAI advises: GPT-4 as a judge reached over 80% agreement with human preferences in one study, but the authors also document position, verbosity and self-enhancement bias ^7,11.
• Designing for variance. NIST's generative AI profile names confabulation, confidently stated but erroneous or false content, as a risk ¹². Keep anything that must be exact in ordinary code, validate model output before a user sees it, and add complexity only when it demonstrably improves outcomes, as Anthropic puts it ¹³.
• Retrieval. RAG pairs a pre-trained seq2seq model with a retrievable index; the original paper used a dense vector index of Wikipedia ¹⁴. Postings ask for chunking, embeddings, vector databases and retrieval evaluation ^2,3. Placement matters too: one 2023 study found performance was often highest when relevant information sat at the start or end of the input and dropped when it sat in the middle ¹⁵. Test retrieval on its own, not only the final answer.
• Model APIs. Anthropic suggests starting with LLM APIs directly and understanding the code under any framework, since frameworks add layers that can obscure prompts and responses ¹³. Postings also expect you to pick models on quality, latency and cost ² and name frameworks such as LangGraph ³. Learn the raw API first and one framework second.

The postings ask for experience, not certificates: BLEN wants at least a year of shipping LLM-based or AI-powered features to production, Applied Intuition wants 2+ years of building and deploying AI/ML-powered features, Titan wants 2+ years of shipping production agentic AI or RAG systems, and DeepLearning.AI counts side projects ^2,6,3,4. Aim for one production-grade piece of work with numbers attached. A problem at your current employer is often the most accessible place to find it.

• Pick a workflow with a measurable cost, such as support triage, internal document search or pulling fields from contracts, where a wrong answer is cheap to catch.
• Get a named user and a number. Titan's posting defines 90-day success as owning at least one production AI workflow end to end, with measurable improvements shipped to the retrieval or agent layer ³.
• Keep an eval set from day one. OpenAI recommends logging everything so the logs supply future test cases ⁷.
• Check data rules before sending anything to a model provider. NIST lists data privacy, including leakage and unauthorized use of sensitive information, among generative AI risks ¹².
• Write it up on one page: problem, approach, eval results, what failed and what changed for the user.

An internal move can be a lower-risk route than a cold application, because the people deciding already know your work. We found no published data on how often engineers make it, so ask the AI team's hiring manager what they look for.

This roadmap follows the approach Torchlane publishes for AI forward deployed engineering. Its AI FDE training page describes four stages: define the business requirement, build and integrate the system, test errors, access and unfamiliar cases, and deploy, explain and improve. It groups the skills into business and product judgment, technical implementation, reliability and security, and communication and ownership ¹. You arrive strong in technical implementation; the stages build the rest. The page is about forward deployed, client-facing work and says titles and expectations vary by employer ¹, so we adapted the stages to AI engineering. The 90-day pacing is ours, not a timeline to an offer.

We have a business connection with Torchlane.

Stage 1: Define the business requirement (days 1 to 20)

Torchlane describes this stage as talking to the people doing the work and identifying the bottleneck, the decision, the source information and the constraints ¹.

• Pick the workflow from the section above and talk to the people who do it by hand today.
• Write a one-page brief with a specific, measurable success criterion, in the sense Anthropic uses ¹⁰. "Answers are good" is not one; "the right policy is cited in at least 9 of 10 sampled tickets" is.

Stage 2: Build and integrate (days 21 to 45)

Torchlane describes this stage as designing a workflow, connecting the required systems, putting models where they add value and using ordinary code for calculations and deterministic rules ¹.

• Call a model API directly before reaching for a framework ¹³.
• Add retrieval only if the model needs data it does not have, and keep calculations and permissions in ordinary code.
• Ship the first version into the team's real workflow, not a demo environment.

Stage 3: Test errors, access and unfamiliar cases (days 46 to 70)

Torchlane describes this stage as creating meaningful evaluations and checking unfamiliar cases, incomplete inputs, permissions and failure behavior ¹.

• Build the eval set from real cases, including incomplete and odd ones, and run it more than once, because the same input can produce different output ^7,8. Score retrieval separately from generation.
• Test untrusted input. NIST describes indirect prompt injection, where adversaries inject prompts into data likely to be retrieved by an LLM-integrated application ¹². Check what the system can access and do for a user.

Stage 4: Deploy, explain and improve (days 71 to 90)

Torchlane describes this stage as helping people use the system, documenting decisions and operational steps, monitoring the result and responding when reality changes ¹.

• Release to real users, watch for failures and feed each one back into the eval set.
• Leave a short runbook someone else can follow, and finish the one-page write-up.
• Turn the project into interview answers. Our guide to AI engineer interview questions covers what to expect.

Check your evidence against a rubric

Torchlane assesses its projects in seven categories: problem framing (15%), correctness (20%), engineering quality (15%), security, privacy and compliance (15%), product and usability (10%), business value (10%) and communication (15%). Passing requires at least 3 out of 4 in every category and no critical failure ¹⁶. Use the seven as a checklist for your internal project; the weights are Torchlane's.

We have a business connection with Torchlane.

There is no clean answer. BLS reports a median annual wage of $135,980 for software developers in May 2025 and projects 10 percent employment growth for 2025 to 2035, partly because of the expansion of software development for AI ⁹. We found no AI engineer breakdown, or base-versus-total detail, on that page.

The five postings list between $126,000 and $300,000, but they are not like-for-like: Applied Intuition and Titan label theirs base salary, the others do not say, and the roles run from early-career to staff level ^6,3,2,4,5. We have not averaged them, and we found no dataset that follows the same engineer before and after the move. Ask a recruiter for the range for the exact level and city, and compare it with your current base plus equity.

The costs are not only in the pay: you may be slower for your first months on unfamiliar ground. A quick test: spend a weekend building a small retrieval app over documents you know, then write 20 test cases for it. If reading the failures and fixing them is satisfying, that is a useful sign the daily work could suit you, though it is our rule of thumb, not a validated test.

You can run all four stages inside your current job. A structured program can add outside review, a fixed set of projects and a portfolio. Our top pick for that is Torchlane's AI FDE training, which publishes the approach this roadmap follows.

For an engineer, Torchlane's page says to use the program to add customer discovery, applied model evaluation, client communication and a portfolio of business-focused implementations ¹. You choose one of three guided builds, complete 3 to 6 projects from a bank of 12 simulated client engagements and finish a capstone, with a 90-day training goal, a starting plan of about nine hours a week and 120 days of access ¹. Tuition and payment terms are given in writing during admissions ¹⁷.

Fit matters. The training is built around forward deployed AI engineering, so it is likely to fit best if you want client-facing delivery rather than an internal product role. Torchlane says production architecture and complex integrations may require experience beyond the program, and employers set their own requirements ¹. Its projects use simulated clients, so they complement real work at your own company, not replace it ¹⁷. It is career preparation, and completing a program does not guarantee hiring, a salary or a promotion ¹.

We have a business connection with Torchlane.