Thirty-five company sections turn a huge field into a study map
The README covers 35 employers, including frontier labs, large technology companies, AI infrastructure vendors, product companies, and forward-deployed teams. Each section names relevant roles, sketches a publicly reported interview loop, and groups prompts under topics such as model internals, serving, retrieval, agents, system design, coding, evaluation, safety, and multimodal systems.
The introduction names 15 job families, from AI Engineer through LLMOps Engineer. AI engineering now stretches from transformer math to GPU capacity planning and production permissions. A candidate can start with questions shared across companies, then use the employer sections to shift emphasis. NVIDIA and Together AI lean toward inference economics, while Glean emphasizes retrieval and enterprise permissions. The guide becomes a checklist for choosing what to study next, not a set of flashcards to memorize in order.
Ten shared topic groups prevent repeated company questions
The common section uses 10 topic groups instead of copying the same KV-cache, batching, RAG, agent, and evaluation prompts into every employer entry. Each repeated question names companies associated with that topic and links an answer where one exists. This makes recurring themes visible: candidates can see that serving tradeoffs or permission-aware retrieval matter across several roles rather than reading them as one company's quirk.
Across the 35 company sections, some questions have articles or videos, often from Outcome School, while many advanced prompts have no answer underneath them. That can be useful for mock interviews because the candidate must build an explanation. It is less useful if you expect a self-contained course. The README promises answers only where they are available and says more will be added.
What happened when we ran it
Our lab did not run commit b62705c. The repository's primary language is Markdown, which is not one of the supported executable ecosystems in the sandbox, and there is no Dockerfile that supplies a runnable path. We therefore have no install time, build result, test count, dependency footprint, or vulnerability scan to claim.
The repository has 4 top-level entries: the README, license, Git attributes, and assets. There is no application or exercise runner. A passing software test would not establish whether an interview question was really asked, whether an answer is correct, or whether a company's loop still follows the described sequence. Those require editorial sourcing and subject review.
Thirty-five company headings lack question-level provenance
The introduction says the 35-company material comes from publicly reported interview experiences and that nothing is confidential. It also warns that loops change by team, seniority, region, and time. Those are responsible caveats. The weakness is that many company-specific questions and loop summaries do not carry a nearby link to the report behind them. A reader can see the claim but often cannot inspect its date or context.
On October 7, 2026, the practical use was to treat each prompt as plausible preparation material, especially when it tests a durable skill such as estimating KV-cache memory or designing tenant isolation. Do not assume a heading means the named company will ask the exact wording. Recent interview reports, recruiter guidance, and the posted job description should decide your final preparation mix.
A 70B memory estimate is better practice than another definition
Many questions go beyond "What is RAG?" or "Explain attention." They ask you to diagnose doubled p99 latency, budget a 70B-model deployment, prevent unauthorized retrieval, handle tool-call failures, or decide whether a model should reach customers. Company sections add realistic constraints, such as clinical provenance for Abridge, low-latency speech for ElevenLabs, and physical safety for Figure AI and Waymo.
The 35-employer set works well in a study group. One person answers, another challenges the assumptions, and a third checks the math or architecture against primary documentation. The unanswered entries are often the most useful because they expose whether you can structure uncertainty. A linked blog post can fill a knowledge gap, but it cannot replace practicing a live tradeoff under time pressure.
Apache-2.0 content also funnels readers to Outcome School
The README identifies Outcome School as the maintainer, promotes its AI and machine learning program near the top, and links many answers to Outcome School articles, videos, and social posts. The project remains Apache-2.0 licensed and freely readable, but it also functions as a path into the maintainer's educational material. Readers should know that before treating the outbound links as a neutral bibliography.
Across 10 shared topic groups, several answer links address the exact concept in the prompt, such as grouped-query attention, prompt caching, or retrieval evaluation. Still, compare explanations with papers, official system documentation, and your own calculations. For company-specific preparation, source dates and role context matter more than the polish of an answer page.
An October 6 push is fresh, while zero issues limits feedback visibility
GitHub showed 1,685 stars, 167 forks, and 0 open issues or pull requests on October 7, 2026. The last push was October 6, so the material had changed one day before this review. There is no tagged release, and the repository exposes no open feedback queue about disputed questions, missing citations, or incorrect answers.
The 35-company structure saves planning time, but fresh edits do not solve provenance. A release tag would help candidates pin the version they studied, while dated citations beside interview-loop claims would make updates auditable. Until then, use the guide as a broad preparation index. The final burden of validating a prompt and producing a sound answer stays with you.