Jane Street AI Engineer Case Interview: Designing and Shipping a Real‑Time Microstructure Signal

This Jane Street–style, highly interactive case simulates partnering with traders and engineers to design, evaluate, and safely deploy an online model that lowers execution cost while market making across multiple venues. You’ll be given a concrete objective (e.g., reduce short‑term adverse selection and slippage on quoting/hedging flows) and a realistic constraints set: millisecond-to-submillisecond tick data, level-2 order books, venue fees/rebates, queue‑position estimates, inventory signals, and strict p99 latency and determinism requirements. The interview emphasizes Jane Street hallmarks: probing ‘why’ questions, collaborative back‑and‑forth, clear reasoning under uncertainty, comfort with simple-but-correct solutions, and humility when trade‑offs surface. Focus areas: (1) Problem framing and metrics—translate the trading goal into an explicit objective (expected execution cost, price impact, and risk-adjusted utility), pick online/offline metrics, and define success thresholds; (2) Modeling and features—propose lightweight features robust to regime shifts (e.g., imbalance, cancel/replace rates, recent microprice moves, short-horizon volatility), guard against leakage/look‑ahead, discuss calibration and confidence estimates, and justify model class (logistic/linear with monotone transforms, tree models with constraints, or distilled NN with bounded inference); (3) Systems and latency—design an inference path that respects cache/memory layout, precomputation, and batching; reason about p95/p99 latency, tail behavior during macro events, and failure isolation; sketch 10–15 lines of pseudocode/Python for an O(1) rolling feature update and a fast scoring function; (4) Evaluation—outline a backtest that respects queue position and partial fills, avoids replay bias, uses proper cross‑validation by event time, and reports treatment effects with uncertainty (e.g., bootstrap CIs); explain how you’d detect overfitting and multiple‑testing pitfalls; (5) Causality and experimentation—propose a safe canary/ABA rollout on a subset of symbols/venues, power calculations, guardrails (kill switches, exposure caps), and monitoring for drift and model decay; (6) Risk and operations—discuss failure modes (stale features, clock skew, data gaps), alerting, and fallback behavior; (7) Communication—think out loud, ask clarifying questions, and teach back trade‑offs, reflecting Jane Street’s culture of open discussion, training, and teamwork. Expect light whiteboard probability or mental‑math checks (e.g., deriving a simple Poisson/Geometric approximation for ‘fill‑within‑Δt’) woven into the case. The bar is high on clarity, correctness, and practicality under real constraints rather than flashy complexity.