Companies · Mercor

Companies

Mercor

Prep plan for a Mercor SWE onsite. System design, a multi-part coding block, and an algorithms discussion.

1 min read system-design

Software Engineer, Agents Onsite In person

System design

45 minutes · Live discussion, no coding. Draw it.

What's scored

  • Product sense: a small set of requirements and a design that matches them.
  • Data storage and search: real tables, indexes, filters, and ranking.
  • Optimizations you can justify with a cost, not a logo.
  • Technical communication, including failure modes you raise yourself.

Checklist

  • By minute 25: requirements, API, entities, data model, and a high-level drawing.
  • Write two indexes and one SQL statement for the hot path.
  • Name three failure modes before they ask (dependency down, duplicate work, hostile input).
  • Offer one deep dive and let them pick.

Linked topics

Practice

  • Prompt: search and rank experts for a role (talent marketplace page), 45 minutes, timer at 25.
  • Prompt: route labeling tasks to experts, including the SKIP LOCKED claim.
  • Prompt: an eval harness for coding agents, using your sandbox split.
  • Prompt: grade an interview with a rubric, versions, and a human review queue.
  • Failure-mode drill: pick any box and list six ways it breaks, in three minutes.

Coding challenge

60 minutes solo, then 30 minutes review · Multi-part problem on your laptop. Correct first, then faster.

What's scored

  • Fundamentals, especially graphs: traversal, cycles, reachability, complexity.
  • Problem solving and original thinking on later parts.
  • Communication in the review: approach, complexity, tests, what you would change.

Checklist

  • Read all parts. Say the structure out loud in the first five minutes.
  • unordered_map and functions. A class only if the API is the problem. Write C++.
  • Tiny tests: empty, single node, cycle, disconnected.
  • A comment per part with time and space, read out before they ask.
  • A correct partial beats a half-finished optimal on every part.

Linked topics

Practice

  • Ladder: course schedule (cycle), course schedule II (order), parallel courses (longest path in a DAG), recipes from supplies.
  • Union-find: redundant connection, accounts merge.
  • Dijkstra and 'at most K edges' (plain Dijkstra is wrong there).
  • Mock: dependency resolver in four parts (transitive deps, install order, the cycle path, parallel install time plus who must reinstall).

Algorithms

30 minutes · Live reasoning about a messy quantitative problem. Not a memorized puzzle.

What's scored

  • Reasoning rather than retrieval.
  • Taking a hint immediately.
  • Putting a number on the trade-off (calls, dollars, latency, accuracy).
  • Saying where your approximation is wrong, before they do.

Checklist

  • Baseline, cost, better idea, new cost, where it breaks, how you would notice.
  • If they nudge, say the nudge changed the design. Do not defend the first idea.
  • Stay precise when you are unsure. Hand-waving is the negative note.

Linked topics

Practice

  • Dedup a million profiles with at most 50k model calls. Block, then score, then spend the model on the uncertain band. Watch transitive merges.
  • Pick the top 10 of 500 candidates when every score is noisy and you have about 3,000 evaluations.
  • Schedule 2 million grading calls under a token rate limit and a dollar budget.
  • Assign 5,000 tasks to 400 experts with capacity. Greedy versus min-cost flow, and where greedy is wrong.
  • Estimate a pass rate within two points. Paired comparisons on the same tasks.

How you ship, and how you use models

Woven through the day · Bring it up. They will not always ask.

What's scored

  • Evidence you ship quickly, with a number and a time box.
  • Evidence you use coding agents in real work and still own diagnosis and tests.

Checklist

  • 90 seconds: situation, what was at stake, what you did, a number, how the model was used, how long it took.
  • Do not recite. Stop and let them ask.

Linked topics

Practice

  • Say two project stories out loud in under two minutes each.
  • Prepare the tools answer with one habit you actually have.

The day

Several sessions, each scored by a different engineer, so one weak session doesn’t decide it. Ask what tools are allowed in each block. Don’t assume an assistant is allowed.

Three habits, until they are boring: name the structure in the first two minutes, say the cost when you pick an approach, and say what breaks before moving on.

Using coding agents

Have a short, honest answer ready: what I use them for (scaffolding, mapping unfamiliar code, drafting tests) and what I keep for myself (diagnosis, test design, reviewing every diff). One concrete habit beats a list of tools.

Stories

About 90 seconds each. One number I can defend, one line on how I used a model, then stop and let them ask. Keep job titles and dates consistent with the resume they have.

Questions to ask

  • What does the team own end to end, and what is hard about it right now?
  • How do you know an agent pipeline change worked?
  • What does a normal week of shipping look like, and how big is a PR?
  • Where do cost and quality trade off: model choice, number of calls, or human review?

Bring

A laptop with g++ -std=c++20 working offline, a scratch repo, a drawing tool, a charger, and a display adapter.

Further reading

Check yourself

Answer the questions. Multiple choice is scored on this device only.

1. On the 45-minute design, when should the drawing be done?

2. In the algorithms round, a hint means:

3. What do you do in the first five minutes of the coding block?