Interview prep
Cursor Model Routing & Inference Engineer Interview Questions
Cursor's Software Engineer, Model Routing & Inference builds the inference platform behind every AI interaction in the product: a gateway over every provider's API, cross-provider failover, and routing backpressure. The posting asks for high-throughput distributed-systems experience and good calls on cost, latency and reliability. Its published process starts with two or three short technicals and ends with an onsite.

On this page
What does a Cursor Model Routing & Inference engineer do?
The posting gives the team's scope in one sentence: it owns the full inference path, making Cursor's AI faster, more reliable and more cost-effective at a scale few teams in the world get to operate at. Every agent session, every tab completion and every chat message flows through this stack. An engineer on it builds the inference platform that powers every AI interaction in the product.
This is a backend platform job, and the posting never mentions training a model or writing a prompt. It talks about provider API semantics, outages, traffic spikes and GPU utilization.
If the word inference made you think of research, reread the example projects: they describe traffic engineering for model requests.
- Team
- Engineering, on Model Routing & Inference.
- Locations
- New York; San Francisco.
- Employment
- Full-time.
- Shape
- Platform engineer on the inference path: gateway, failover, routing and admission control.
- Compensation
- The posting gives no range.
Cursor lists three example projects, and each one is a system rather than a feature. Read them as three failures the team works to prevent: slow model onboarding, a provider outage that users can see, and a traffic spike that cascades into a provider.
- Building and evolving the inference gateway, a single abstraction over every provider's API semantics, so model onboarding becomes a config change.
- Designing intelligent cross-provider failover so no single provider outage causes user-visible degradation.
- Designing routing backpressure and admission control so traffic spikes do not cascade into providers.
This exact topic is a hands-on Lesson: LLM Inference Fundamentals for the Routing Engineer — about 27 minutes, free to read.
Rather do it than read about it? Run 11 interactive Cursor walkthroughs in a simulated editor. Free, no account needed.
What does this role own, and what does it not?
Unlike the BugbotCursor's automated PR reviewer that posts inline findings and can push fix commits from isolated VMs. Press Enter for the full definition. posting, this one has no list of what you won't own, so read the boundary from what it names: the gateway, failover and admission control. It doesn't name model quality, prompting, the agent harness or the client. The archived Agent Harness posting gave that team the model choices for Auto and new model launches.
- You own
- The inference path: the gateway over every provider, failover between providers, routing backpressure and admission control.
- Adjacent, not yours
- Model quality and prompting; which model a product surface picks; the agent harness; the client.
- Stated bar
- Production systems that handle millions of requests, and good calls in the gray area between reliability, cost, latency and user experience.
Prepare against the gray-area line. The fourth fit bullet asks for good calls when there isn't a single right answer, weighing reliability, cost, latency and user experience. Every example project contains that tension. A system that fails over instantly can stampede onto the backup provider, and one that admits every request can cascade.
Where does Cursor Router fit? Cursor's docs call it the system that routes Auto requests and limit it to Teams and Enterprise plans, though Cursor's help-center pricing page says it "will launch" on those plans. In Balance or Intelligence mode, it classifies every agent request by task type and complexity and sends it to a model. The posting doesn't mention Cursor Router, so don't assume which team owns it. Do know what happens to a request after the routing decision, because every agent session flows through this team's stack.
The gateway's stated purpose is that model onboarding becomes a config change. That is a strong claim about abstraction: every provider's streaming format, tool-call schema, rate-limit semantics and error taxonomy has to fit behind one interface. Prepare to say where that abstraction leaks.
What does the Model Routing & Inference interview assess?
Unlike most Cursor postings, this one ends with a note on applying. For candidates who seem to fit, the note sets out two or three short technicals, and after them "an onsite in our office, where you'll work on a small project, discuss ideas, and meet the team." The note is silent on what the stages contain.
Deep experience building high-throughput, low-latency distributed systems, with inference serving, traffic routing and real-time data pipelines named as the specific flavors.
Comfort reasoning about cost and performance trade-offs at scale: GPU utilization, provider economics and capacity planning are the three examples the posting gives.
Strong software engineering fundamentals, a record of shipping production systems that handle millions of requests, and good calls in the gray area between reliability, cost, latency and user experience.
Two of the three are standard senior-backend criteria. The middle one is unusual, and we think it is the one most likely to catch a strong systems candidate with no inference exposure. Provider economics means knowing that a token's price varies by provider, by cache hits, and by fast or standard speed, so a routing decision is also a spend decision. You don't need Cursor's numbers, but you should show you have thought in those units before.
The note doesn't say what the onsite project is, how long it runs or what tooling is allowed. Practice building something small that works, against a clock, in your strongest language.
What interview questions should I expect?
Read the table left to right: a line from the posting, the question it suggests, and how strong and weak answers differ. We wrote the questions, because Cursor hasn't published any for this role.
- From the posting
- A single abstraction over every provider's API semantics
- Expect to be asked to...
- Design the gateway interface, then say which provider behaviors refuse to fit behind it.
- Strong answer vs weak
- Strong: names a concrete leak (streaming, tool calls, errors) and how it is contained. Weak: a clean interface with no leaks.
- From the posting
- Model onboarding becomes a config change
- Expect to be asked to...
- Walk through adding a new provider end to end and identify the step that is not config.
- Strong answer vs weak
- Strong: an honest non-config step, such as validating semantics. Weak: "it's just YAML."
- From the posting
- Cross-provider failover with no user-visible degradation
- Expect to be asked to...
- Decide when to fail over, how to detect the outage, and how to avoid stampeding the backup.
- Strong answer vs weak
- Strong: a detection signal, a budget and a ramp. Weak: "retry on the other provider."
- From the posting
- Routing backpressure and admission control
- Expect to be asked to...
- Handle a traffic spike so it does not cascade into providers; say who gets rejected first and why.
- Strong answer vs weak
- Strong: a priority order tied to user experience (tab vs chat vs agent). Weak: a global rate limit.
- From the posting
- GPU utilization, provider economics, capacity planning
- Expect to be asked to...
- Estimate the cost of a routing change and the capacity you would reserve for a launch.
- Strong answer vs weak
- Strong: a back-of-envelope with stated assumptions. Weak: "we'd monitor it."
- From the posting
- Millions of requests in production
- Expect to be asked to...
- Tell the story of an incident on a system you ran at scale and what changed afterwards.
- Strong answer vs weak
- Strong: a change to the system, with the number that showed it worked. Weak: a postmortem with action items nobody owned.
- From the posting
- Good calls in the gray area
- Expect to be asked to...
- Pick between a faster and a more reliable configuration for a described workload, and defend it.
- Strong answer vs weak
- Strong: commits to one and names the condition that would flip it. Weak: "it depends" with no condition.
| From the posting | Expect to be asked to... | Strong answer vs weak |
|---|---|---|
| A single abstraction over every provider's API semantics | Design the gateway interface, then say which provider behaviors refuse to fit behind it. | Strong: names a concrete leak (streaming, tool calls, errors) and how it is contained. Weak: a clean interface with no leaks. |
| Model onboarding becomes a config change | Walk through adding a new provider end to end and identify the step that is not config. | Strong: an honest non-config step, such as validating semantics. Weak: "it's just YAML." |
| Cross-provider failover with no user-visible degradation | Decide when to fail over, how to detect the outage, and how to avoid stampeding the backup. | Strong: a detection signal, a budget and a ramp. Weak: "retry on the other provider." |
| Routing backpressure and admission control | Handle a traffic spike so it does not cascade into providers; say who gets rejected first and why. | Strong: a priority order tied to user experience (tab vs chat vs agent). Weak: a global rate limit. |
| GPU utilization, provider economics, capacity planning | Estimate the cost of a routing change and the capacity you would reserve for a launch. | Strong: a back-of-envelope with stated assumptions. Weak: "we'd monitor it." |
| Millions of requests in production | Tell the story of an incident on a system you ran at scale and what changed afterwards. | Strong: a change to the system, with the number that showed it worked. Weak: a postmortem with action items nobody owned. |
| Good calls in the gray area | Pick between a faster and a more reliable configuration for a described workload, and defend it. | Strong: commits to one and names the condition that would flip it. Weak: "it depends" with no condition. |
Questions we wrote from the three example projects and four fit bullets.
Rehearse the last row out loud. "It depends" is a weak answer on its own. A strong one names what the decision depends on and what you would do in each case.
How do I prepare for the Cursor Model Routing & Inference interview?
Build a small version of the system the posting describes. A gateway that fronts two model providers, streams tokens, fails over when one returns errors and sheds load when a queue grows is a weekend project. Once you have built one, your answers describe things you watched happen instead of theory.
- 1Write a two-provider gateway that normalizes streaming and error semantics behind one interface. Keep a list of every place the two providers refused to look the same.
- 2Add failover with a budget. Detect provider failure from a rolling error rate, fail over with a ramp rather than a cliff, and write down the condition that moves traffic back.
- 3Add admission control with priorities. Decide which request class you would reject first under a spike and why; Cursor's own surfaces (tab completion, chat, agent sessions) are a good set to reason about.
- 4Do the economics on paper. Price a day of traffic through your gateway under two routing policies, using public provider list prices, and be ready to explain the assumptions.
- 5Prepare one incident story from a system you ran at scale, told as detection, decision, fix and the measurement that proved the fix.
Don't skip capacity planning because it sounds like a manager's job. The posting lists it with GPU utilization and provider economics as something you reason about, so a launch that doubles agent traffic becomes your problem early. Have an answer for how much headroom you would reserve for a model launch, and what you would do if you were wrong.
Our ML & Inference practice track, one of the free interview-prep paths, was written around this posting and drills gateway design, failover and backpressure. The Software Engineer, Infrastructure track covers the cloud and cluster foundation underneath. Both tracks include practice rounds that this posting doesn't describe. For the real process, rely on the posting's closing note alone.
Each day, take one routing or reliability decision from a system you know and write the trade-off in one sentence with a number in it. For example, name the latency you would accept for the reliability you would gain, or the spend you would add for the failover you would get. A week of those sentences gives you the gray-area judgment the posting asks for, in writing.
Frequently asked questions
What does the posting say about the interview process?
Only the outline. Its closing note gives the order of steps and little else: a first round of two or three short technicals, and after that an office onsite with a small project on the agenda. It gives no detail on what any stage contains.
Is the Model Routing & Inference role remote?
No remote option is listed. The posting names New York and San Francisco, full-time on the Engineering team, and ends with an onsite stage in the office. It publishes no compensation range.
Is ML research experience required for this role?
It isn't in the fit criteria. Those ask for high-throughput, low-latency distributed systems, cost and performance reasoning at scale, strong engineering fundamentals, and judgment on reliability, cost, latency and user experience. The posting names inference serving as one kind of systems experience, not as research.
What does the Model Routing & Inference team own?
The posting says the team owns the full inference path: the inference gateway over every provider's API, cross-provider failover, and routing backpressure and admission control. Every agent session, tab completion and chat message in Cursor flows through it.
Sources & last verified
- Cursor - Software Engineer, Model Routing & Inference (careers)
- Cursor - Careers
- Cursor Docs - Cursor Router
- Cursor Help - Pricing and plans
Cursor ships frequently. Last updated October 28, 2026.