Luma is a small-to-mid-size company, currently around 200-300 people, split mainly into research and product. The process is fast and also a bit chaotic: recruiter, take-home, onsite, offer. If the recruiter doesn't get back to you about the next step, you can follow up yourself. TC is very high, they mainly want to be in the same range as OpenAI and Anthropic, but most of it is Series C paper money, and there's still a while to go before an IPO.
Recruiter call - motivation, why Luma, experience with AI development tools and workflows, what specific interests in dev.
OA / take-home - 8-12 hours, pick one of five prompts, with frontend, backend, and infra all covered. Judging from the interview reports on Glassdoor, Reddit, and Blind, the question bank should be more than just 5 questions. The recruiter email has a curl script that directly downloads the prompt's README and the submission script. Your Cursor, Codex, and Claude Code chat logs all get uploaded, so what they're testing should be agentic best practices, 0-to-1 product development scoping, and UX/dev experience design.
- build scalable TTS system
- managing prompt & model behavior in prod
- where and when to eat with friends
- collaborative document editing with people and agents
- reverse engineer undocumented APIs from another app and rebuild app to improve UX
Onsite - the office is in Redwood City and it's already getting a bit full. The recruiter said the company has already started looking for expansion space in SF.
Take-home product review - this round is mainly a discussion of the take-home's trade-offs, scaling limitations, customer needs, and UI/UX. The questions are very detailed; it seems like there's an AI judge on the backend that reads your take-home, catches all the weaknesses, and hands them to the interviewer to quiz you on. Think of the one day of work you did as a prototype, and after that they add other customer requirements and discuss integration, estimation, and roadmap.
Leadership - this round was with the CTO, who is really smart and knows everything. He also talks really fast, so be mentally prepared. We mainly discussed potential projects, interests, and team.
System design - in the prep call beforehand, the recruiter said the question would be something in the rideshare / payment processor category. I got build Uber, which is arguably one of the most complex SDs out there. The interviewer used to be a backend systems engineer at a big company doing pretty much this, so you can't fake anything past him. The Hello Interview design isn't enough; also look at the 99.99% availability design from another prep site plus one more write-up, and compare all three. Honestly every one of them has weaknesses, and if you find the trade-offs you can bring them up proactively. The ChatGPT and Claude versions both have a lot of holes too.
- requirements and scope clarification
- entities, database design, and API endpoints
- system diagram
- idempotent tx, atomic writes, double charge
- database design, peak throughput, scaling read/writes, sharding
- distributed locks
- geospatial search
- tx logging, error/crash recovery, replay
- audit trail and reconciliation
- workflow tooling and durable execution
- cancellation handling
- client/server state updates: SSE vs. WebSocket vs. push notification
- system scaling and weak points
My overall feeling is that the people inside are really smart, but product-market fit is still a bit lacking. Research has already started working on things like robotics and real-world models, while the product side is still building gen AI video and image tools to sell to big media and advertising customers.
Discussion
Loading comments…