Engineering

Apply for AI-Augmented Software Developer (New Grad)

Takes about 15 minutes. We read every application ourselves.

Step 1 of 8About you

1About you
2Questions
3AI Fundamentals
4Reading AI Code
5Judgment & Ethics
6Agentic Systems
7Show Us
8Review

About you

Full name *Email *Résumé *PDF, DOC or DOCX

A few more questions

Portfolio, GitHub, or something you’ve shipped *A link to something real you built — however small.Which AI-augmented development program did you complete?Bootcamp, course, or a self-directed path — tell us about it.LinkedIn

Screening

AI Fundamentals

A hands-on screen — no trivia. We care how you work WITH AI tools: how you verify, when you don’t trust them, and whether you understand the code that ships. A few questions are short-answer and read by a human.

1. You ask an AI assistant to "add pagination to this Supabase query" and it returns working-looking code. Before committing, which single action gives you the MOST confidence the code is correct?

Ask the same assistant “are you sure this is correct?” and trust its answerRun it against real/seeded data and confirm the returned rows, counts, and page boundaries match what you expectCheck that the code compiles / has no TypeScript errorsConfirm the code matches the style of the surrounding file

2. Which of these make an AI coding prompt MORE likely to produce correct, usable output? (Select all that apply.)

Select all that apply.

Pasting the actual relevant types / schema / function signatures it must conform toStating explicit constraints and the runtime (e.g. “Astro SSR endpoint, Node 20, no external deps”)Asking for the whole feature in one giant prompt so it has maximum freedomTelling it what you already tried and what failedAdding “make it production-ready and secure” and assuming that covers your requirements

3. An AI assistant confidently writes supabase.from("patients").select("*").where("id", patientId) using the official supabase-js client. What is the MOST likely problem?

.where() is not a method on the Supabase query builder — it hallucinated a Knex/SQL-style API; the correct call is .eq("id", patientId)select("*") is invalid syntax and must be select()The query is fine; it just needs to be awaitedfrom("patients") should be table("patients")

Screening

Reading AI Code

4. An AI generated this Astro API endpoint. What is the MOST serious problem? export async function GET({ url }) { const id = url.searchParams.get("id"); const { data } = await supabase .from("patients") .select("id, name, dob, ssn, notes") .eq("id", id); return new Response(JSON.stringify(data), { status: 200 }); }

It returns data even when error is set, so failures are silently ignoredThere is no auth/authorization check — any caller can fetch any patient’s PHI (incl. SSN) by guessing an id (broken access control / IDOR)url.searchParams.get can return null, which is a type errorJSON.stringify is unnecessary because Response serializes automatically

5. An AI wrote this “fetch with retry” helper. Which statements are TRUE? (Select all that apply.) async function fetchWithRetry(url, retries = 3) { for (let i = 0; i < retries; i++) { const res = await fetch(url); if (res.ok) return res.json(); } }

Select all that apply.

On a non-ok response it retries immediately with no backoff, which can hammer the serverIf all attempts fail it returns undefined, and callers awaiting it won’t get an error to catchA network throw (e.g. DNS failure) rejects on the first attempt — it is not actually retriedIt correctly distinguishes retryable (5xx) from non-retryable (4xx) responsesres.json() is guaranteed to succeed whenever res.ok is true

6. You asked for a function to check whether an appointment time is in the future. The AI produced: function isFuture(dateStr) { const d = new Date(dateStr); return d > new Date(); } Best assessment?

Correct and complete; ship itIt works for valid input but silently misbehaves on invalid dateStr — new Date("garbage") is Invalid Date, comparisons are false, so bad input is treated as “not future” with no error. Needs input validationIt’s wrong because you can’t compare Date objects with > in JavaScriptIt’s wrong because it ignores time zones, making the comparison meaningless

Screening

Judgment & Ethics

7. For which task is it LEAST appropriate to accept AI-generated code without deep verification?

A CSS tweak to center a marketing hero sectionA regex to loosely highlight phone numbers in non-critical UI textThe logic that decides which patients are eligible to be prescribed a specific medicationRenaming a variable across a file

8. You want to paste some real production data into a public AI chat tool to debug a query. Which are appropriate precautions or correct judgments? (Select all that apply.)

Select all that apply.

Don’t paste real PHI/PII into a third-party tool that isn’t covered by a BAA / approved for that dataReproduce the bug with synthetic/anonymized data or just the schema + query shape insteadIt’s fine as long as you delete the chat afterwardRedacting only the name but leaving DOB + zip + diagnosis is sufficient de-identificationPrefer an approved/enterprise AI tool with a data-processing agreement for anything touching sensitive systems

9. In a US telehealth product, “PHI” most accurately refers to:

Any data stored in the patients database tableIndividually identifiable health information created or held by a covered entity, tied to a person’s health, care, or payment for careOnly Social Security numbers and credit card numbersAny personal data — identical in meaning to “PII”

10. Describe a specific time an AI tool was CONFIDENTLY WRONG in a way you caught. What tipped you off, and how did you confirm and fix it? (If you can’t recall one, be honest and describe how you’d expect to catch such a case.)

Screening

Agentic Systems

11. You’re designing an agent that can call a refundPayment(orderId, amount) tool. Which guardrail is MOST important before letting it run unattended on real orders?

Give the model a very detailed system prompt telling it to be carefulEnforce hard limits in the tool implementation itself — validate amount ≤ order total, verify ownership, require approval above a threshold — so a bad model decision can’t cause an unbounded refundUse a larger, smarter model so it makes fewer mistakesLog every action so you can review refunds after they happen

12. Which are good practices when designing a tool for an LLM agent to call? (Select all that apply.)

Select all that apply.

Clear name/description and typed, validated parameters so bad args are rejectedReturn structured, informative errors the agent can act on rather than an opaque stack traceMake each tool do as many things as possible so the agent needs fewer toolsMake destructive/irreversible actions require confirmation or a separate approval stepKeep tool outputs concise and relevant rather than dumping huge raw payloads into context

13. You built an agent that summarizes patient intake forms. Your manager asks, “How do you know it’s actually good?” Strongest answer?

“I tried it on a few forms and the summaries looked good to me”“I built an eval set of representative intake forms with expected outputs/graded criteria, run the agent against it, and track accuracy/failure rates as I change prompts or models”“It uses the latest, most capable model, so it should be reliable”“It hasn’t produced a wrong summary in the demo yet”

14. An agent loop calls tools until the task is “done.” Most important safeguard against it running forever / burning cost?

A maximum step/iteration count (and/or token/time budget) after which it stops and reports failureA more detailed prompt telling it not to loopFaster hardware so each loop is cheaperRetrying the whole task from scratch if it takes too long

Screening

Show Us

15. Two rows exist in a users table. This runs against Postgres: SELECT email FROM users WHERE last_login > NOW() - INTERVAL '30 days'; A user has last_login = NULL (never logged in). Are they returned?

Yes — NULL is treated as very old, so the condition is trueNo — any comparison with NULL evaluates to NULL (not true), so the row is excludedYes — NULL compares as 0 / epochIt throws an error because you can’t compare NULL

16. Row Level Security is enabled on patients so users see only their own rows. Your server-side code uses the SERVICE ROLE key. What happens when it queries patients?

RLS still restricts it to the current user’s rowsThe service role key BYPASSES RLS, so it can read all rows — access control must be enforced by your own server codeThe query fails because RLS blocks all access without a logged-in userRLS only applies to writes, not reads

17. An AI just generated a 120-line file that makes your feature work, but you don’t fully understand ~20 lines of it. It’s Friday and the feature is due. What do you actually do, and why? (Be honest about tradeoffs.)

18. A teammate opens a PR that’s entirely AI-generated and clearly untested — it has an obvious bug and a hardcoded secret. Write the PR comment you’d actually leave. (We’re evaluating tone AND substance.)

19. Briefly: describe one automation or internal tool you’d want to build for a small telehealth company, and one guardrail you’d put on it. (2–4 sentences.)

Almost there

Review & submit

By applying you agree to let Rejuve store your details for this role.