How I’d prep for OpenAI’s Full Stack Software Engineer, ChatGPT Finances role in 30 days
Updated Nov 16, 2026
The job
Full Stack Software Engineer, ChatGPT Finances, OpenAI
San Francisco (hybrid) · $347K – $385K, plus equity · Posted Nov 13, 2026
View the posting →What this team actually does
ChatGPT Finances lets you connect your bank, card and investment accounts to ChatGPT, see all your money in one dashboard, and ask questions about it. The posting says the team builds “account connectivity, data ingestion, dashboards, personalized insights, and conversational experiences.”
The product is already live. OpenAI launched it in May 2026 as a preview for ChatGPT Pro subscribers in the US, on the web and iOS. Accounts connect through Plaid, which covers over 12,000 financial institutions. Access is read-only, and the dashboard shows portfolio performance, spending, subscriptions and upcoming payments. OpenAI built the team partly through two acqui-hires of personal finance startups: Roi in October 2025 and Hiro Finance in April 2026.
The hard parts are in the posting too: ingesting and normalizing financial data, handling “synchronization, data freshness, partial failures, permissions, and user consent”, and answering money questions in a way that is “grounded, transparent, and useful.” It’s a full stack role, so you own all of it, from the React screen to the database.
| Area | What to know |
|---|---|
| Account linking | Plaid (read-only). OpenAI said Intuit is coming next |
| Backend | Python, Go or Node.js (the posting). OpenAI’s other product postings say Python with FastAPI. I’d pick Python |
| Frontend | React and TypeScript on the web; Swift (iOS) or Kotlin (Android) for mobile |
| Database | Postgres. ChatGPT runs on one primary Postgres database with about 50 read replicas |
| AI layer | OpenAI models answering questions from the user’s own data (tool calls, grounded answers) |
| Around it | Auth and permissions, observability, experiments, and privacy for very sensitive data |
How I found the stack (do this for any job)
- Read the posting closely. It lists “Python, Go, Node.js” for the backend and “React / Typescript, or Swift or Kotlin” for the frontend. That’s a lot of options, and it doesn’t say how they connect to your bank.
- Read the launch coverage. When OpenAI launched ChatGPT Finances, TechCrunch reported that it “has partnered with the financial connection service Plaid to manage the account connections.” That’s the integration this team owns.
- Read the company’s other postings for similar teams. The Ads Manager posting asks for “Python / FastAPI and TypeScript / React”. The Applied Emerging Talent posting says “we use Python” and asks for Postgres. The ChatGPT Shopping posting, a sister consumer team, lists React, TypeScript and Next.js. A Developer Productivity posting names “Postgres, Cosmos DB, Kafka, Python, and FastAPI”.
- Read the engineering blog. OpenAI’s post Scaling PostgreSQL to power 800 million ChatGPT users explains how ChatGPT runs on one primary Postgres database with about 50 read replicas.
- Look up the people. OpenAI bought two personal finance startups to build this team, so search LinkedIn for the engineers who came from them and read what they worked on. If you’re feeling bold, send one a short, specific message about the role.
So: Plaid for the bank connections, Python (FastAPI) and Postgres on the backend, React and TypeScript on the front end, and OpenAI’s own API for the answers. That’s where I’d spend the 30 days.
Understand the product first
- If you have ChatGPT Pro (US), use the product. Connect an account, look at the dashboard, and ask it a dozen questions about your spending. Note what’s fast, what’s slow, and where it tells you which data it used.
- If you don’t, read the launch coverage and study the screenshots: the account list with each bank marked “Synced”, the spending dashboard, and the chat that builds a plan from your real numbers.
- Run Plaid’s quickstart in sandbox mode with the test credentials from Plaid’s docs. Link a fake bank and watch which calls happen: link token, public token, access token.
- Then call /transactions/sync yourself. Notice what comes back: “added”, “modified” and “removed” lists, plus a cursor. Fire a sandbox webhook and see what Plaid sends you.
Then ask yourself the questions this team answers every day: how fresh is the data when someone asks a question? What happens when one of their five banks is down? What if the same webhook arrives twice? How does the model know a number is real and not made up? What happens to the data when someone disconnects?
System design questions
The first two are what candidates keep reporting from OpenAI’s system design rounds (see Hello Interview’s OpenAI guide and interviewing.io). The rest are what I’d expect for this team.
- Design a webhook delivery platform with retries over a 24-hour window. What happens when a receiver is down for an hour? For a day?
- Design a notification system (push, email, in-app). How do you avoid sending the same alert twice, and respect each user’s settings?
- Design ChatGPT Finances: link accounts at 12,000 institutions, keep transactions fresh, and show a dashboard that loads fast even when banks are slow.
- You receive Plaid webhooks. They can arrive twice, out of order, or not at all. Design the receiving side so every transaction is saved exactly once.
- A user has 10 linked accounts and one bank is down. What does the dashboard show, and what does the chat say?
- “How much did I spend on food last month?” Design the path from the question to a grounded answer that cites the transactions it used. How do you stop the model from making up numbers?
- Design ChatGPT itself. Why stream the answer with server-sent events instead of WebSockets? How do you keep a very long conversation fast in the browser?
- A user disconnects their accounts. Design the deletion so their data is gone everywhere within 30 days, including caches and backups.
Concepts they’ll probe
- Webhooks: at-least-once delivery, retries with exponential backoff and jitter, dead-letter queues, signing and verifying payloads, and returning 200 fast before doing the work.
- Idempotency: idempotency keys, deduplicating by a stable ID, and Postgres upserts (
INSERT … ON CONFLICT). - Queues and workers: what a worker does after a crash, and whether a job can run twice.
- Sync: cursors, added/modified/removed changes, pending vs posted transactions, and how fresh is fresh enough.
- Partial failures: timeouts, retries, circuit breakers, and showing partial data honestly.
- Postgres at scale: indexes, transactions, read replicas and replication lag, connection pooling.
- Auth and consent: OAuth-style account linking, storing access tokens encrypted, read-only scopes, revoking access and deleting data.
- Grounding AI answers: tool (function) calling, structured outputs, citing the rows used, and evaluating answers for accuracy.
- Streaming: server-sent events vs WebSockets, and rendering streamed text in React.
- Money basics: store amounts as integer cents with a currency, round carefully, and categorize consistently.
- Product engineering: loading, empty and error states, accessibility, experiments and product metrics.
Unlock the rest
- The week-by-week 30-day plan, with daily practice on Devmaxx
- The exact book chapter and docs for each week
- The project: a personal finance copilot with Plaid, FastAPI, Postgres, a React dashboard and a chat that cites your real transactions
- The senior bar: what OpenAI expects at this pay level, and the project deep-dive round
- The resource list with every link
Get the full guide, free
Enter your email and the guide opens right here. You’ll also get the occasional Devmaxx email. Unsubscribe anytime.