10 Citadel Software Engineer (New Grad) Interview Questions (2026)

Outline

STAR. Pick a real bad decision. Walk through your reasoning at the time, why it was wrong in retrospect, the consequences, how you fixed it, what you do differently now. Citadel values intellectual humility — admitting mistakes with calibrated reflection scores high. Avoid blaming others.

Outline

Specific reasons: tight feedback loop between engineering and trading (your code's latency directly impacts P&L), challenging systems work (microsecond latency, terabyte-scale data), focused engineering culture (small teams, deep technical ownership), strong comp. Avoid 'high pay' as the lead. Reference Citadel's specific engineering challenges if you've researched them.

Outline

Two heaps: max-heap for lower half, min-heap for upper half. Maintain invariant: |size_max - size_min| ≤ 1. Insert: add to one based on value vs current median, rebalance. Median: if equal sizes, average tops; else top of larger. O(log n) insert, O(1) median. Walk through example. Code cleanly.

Outline

Multiple approaches: (1) Sort, return arr[n-k], O(n log n). (2) Min-heap of size k: iterate array, push to heap, pop if size>k. Heap top is answer. O(n log k). (3) Quickselect: partition around random pivot, recurse on side containing kth. O(n) average. Walk through heap solution first; mention quickselect for the optimal-time follow-up.

Outline

DFS with memoization. For each cell, compute longest path starting there: 1 + max over neighbors with larger value. Cache results. O(m·n) time, O(m·n) space for memo. Walk through example. Mention topological sort alternative.

Outline

Floyd's tortoise and hare. Two pointers: slow moves 1 step, fast moves 2. If they meet, cycle exists. If fast reaches null, no cycle. O(n) time, O(1) space. To find start of cycle: after meeting, restart one pointer at head; move both 1 step at a time; they meet at cycle start. Walk through why this works (modular arithmetic).

Outline

User process sees virtual addresses. CPU's MMU translates virtual → physical via page tables. Translation involves multiple levels (e.g., x86-64 has 4-5 levels). TLB caches recent translations to avoid walking page tables on every access. Page fault triggers OS to load from disk or kill process. Walk through the full path from C pointer dereference to RAM bytes. Discuss why this matters for low-latency code (TLB misses are costly).

Outline

Mutex: binary lock for mutual exclusion; ownership tied to thread that locked it. Semaphore: counter with wait/signal operations, no ownership. Semaphore with count=1 is similar to mutex but lacks ownership semantics — any thread can signal it. Use mutex for critical sections; use semaphore for resource counting (e.g., connection pool). Discuss priority inversion and how it relates.

Outline

Two main algorithms: (1) Token bucket — refill at rate N tokens/sec, consume 1 per request, reject if empty. (2) Sliding window log — track timestamps, reject if N already in last second. Discuss tradeoffs: token bucket allows bursts, sliding window is strict. Implementation: Redis (atomic INCR with TTL) for distributed; per-instance memory for single-server. Discuss handling thundering herd at second boundaries.

Outline

Multi-layered: (1) Connection management: epoll/io_uring on Linux for event-driven I/O; CPU affinity per core. (2) Memory: lock-free data structures, NUMA-aware allocation, huge pages. (3) Network: kernel-bypass (DPDK) for ultra-low latency, multicast for fan-out. (4) Hardware: bare metal preferred, RDMA NICs. Discuss the order of magnitude work each layer addresses. Citadel cares about low-latency design intuition.