OpenAI SWE Interview Experience – full loop breakdown

Went through OpenAI's loop for a platform SWE (Software Engineer) role. Two weeks start to finish which honestly caught me off guard.

Recruiter call was 20 min, nothing technical, mostly "why OpenAI" and what kind of infra work I care about.

Take Home :

48 hour window to build a Webhook Delivery System. Register endpoints, receive events, deliver reliably, retries with backoff, dead letter queue for permanently failed stuff, and an API(Application Programming Interface) to check status. I did it in Python with FastAPI (Fast framework for building APIs in Python) and SQLite(Structured Query Language). Spent about 6 hours which felt like a lot for a take home but they said they care more about clean code and tests than feature completeness so I leaned into that.

The retry mechanism was the interesting part, separate worker process polling for pending deliveries, exponential backoff, circuit breaking after 10 consecutive failures. Looking back my circuit breaker threshold was probably too aggressive but nobody brought it up so maybe it was fine?

Technical Deep Dive :

The senior engineer reviewed my take home live and this ended up being the best round by far. The first 20 min was walking through decisions (why SQLite, what I'd swap for prod). Then she had me extend the system live, HMAC(Hash-based Message Authentication Code) signature verification and event type filtering.

She caught a bug I completely missed, if the worker crashes mid-delivery the event gets stuck as in-progress forever and never retries. We worked through a lease-based approach together where deliveries auto-requeue if not completed in time. Genuinely learned something from that.

System Design :

This one hurt… Design an in-memory database with basic SQL (CREATE TABLE, INSERT, SELECT with WHERE, JOINs). Went column-oriented since the follow ups were heading toward analytical queries.

JOINs is where things slowed down, starting with nested loop joins, he immediately asked me to do better. Talked through hash join vs sort-merge join. Then he asked about adding transactions with ACID (Atomicity, Consistency, Isolation, and Durability) guarantees and I described WAL(Write-Ahead Logging) plus MVCC but I was definitely getting hand-wavy by that point. He just kept going deeper on every answer, like every response I gave opened two more questions. 60 minutes felt like maybe 20.

Behavioral :

Engineering manager. Technical disagreements, failed projects, prioritization. He asked about a time I pushed back on a technical decision for ethical reasons and I talked about a logging system at my last job that was capturing way more user data than necessary.