Take-Home Assignments

The take-home is the practical exam of AI engineering hiring. Instead of an algorithm puzzle on a whiteboard, you get a small product brief and a few days to ship it: a RAG assistant, an agent with a tool or two, or a chatbot grounded in a FAQ. The AISL material covers this round from both sides: the AI Engineering Field Guide measures what companies actually assign and how they grade it, and an AISL workshop solves a real assignment live, from scoping to submission.

What the market actually assigns

The numbers come from the field guide's interview research — 51 companies with disclosed hiring processes plus 100+ public GitHub repositories of candidate submissions (home assignments):

• 17 of the 51 companies (33%) use a take-home or asynchronous assignment, and 5 more use paid work trials instead.
• The typical format: a 2-7 day deadline for what is meant to be 2-4 hours of work, submitted as code plus a writeup, then defended in a 45-90 minute walkthrough interview. Companies grade decision-making and clarity, not cleverness. The assignments call it "not code golf".
• The task mix across submissions splits into five buckets. RAG systems lead at 40%+, with document upload, vector search, citations, and "I don't know" fallbacks. Agents and tool calling follow at 30%+, conversational AI at 20%+, document processing at 15%, and LLM-as-judge evaluation at 10%+.
• AI tools are effectively allowed: not one company in the dataset bans them for take-homes — the bans apply to live interviews. The assignment tests what you ship with the tools, which is the same stance the AI Coding Tools material takes on daily work.

How submissions are graded

Explicit rubrics recur across the assignments in the corpus. They grade functional correctness, code quality and architecture, and evaluation methodology — whether you build an eval harness and measure quality systematically. They also grade production readiness (caching, monitoring, cost, PII handling) and documentation that explains design decisions and trade-offs rather than restating the README. Performance targets appear as literal numbers: under 2 seconds p95 latency, 100+ requests per second, 40%+ cache hit rates. Testing is sometimes mandatory, with coverage targets around 80%. Some assignments ship weighted rubrics, like 30% functionality, 30% challenge completion, 25% context engineering, and 15% code quality.

What gets an offer

The field guide's preparation advice, distilled from candidate reports and the submission corpus:

• Start with evaluation — build the eval harness before the main logic. YC startups report it as what separates the candidates who get hired: "red flag if candidate doesn't start with evals".
• Ask clarifying questions before starting, and double your time estimate.
• Document decisions and trade-offs, not just code.
• Record a video walkthrough of the submission.
• Make it configurable: one engineer's PDF-summarizer CLI with a config file for models and chunking strategies turned into two competing offers within 72 hours.

The most common failure mirrors that advice: submissions ignore evaluation and testing of AI outputs.

There is also a shadow side: some AI startups use take-homes as unpaid product ideation. The field guide's guidance: decline assignments that constitute a deployable product, and refuse take-homes sent before any human conversation (take-home assignments).

A real assignment, solved live

Solving an AI Engineer Home Assignment: Investment Coach Bot is a real take-home done the way the market data says to do it, in a two-hour session with a coding agent typing. The assignment asks for an investment coaching bot that stays on the educational side of the line. That is the same archetype the field guide's assignment corpus documents: a Telegram bot for investment coaching with safety filtering.

The session walks the full arc:

• Scope first, safety boundary first: no buy/sell/hold recommendations, no position sizing, no price predictions, written into the agent's instructions before any feature work.
• Build a CLI agent with PydanticAI, pivoting from a paid market-data API to free SEC EDGAR endpoints when the free tier blocks the session.
• Cull the pointless tests the coding agent generated, write test guidelines, then real tests against them.
• Add an evaluation harness with scenarios and an LLM judge.
• Wrap the agent as a Telegram bot.

The prompts are included verbatim, so the workflow repeats with any agent. The discipline on display, evaluation before polish, is exactly what the grading data says gets offers. Evaluation and Monitoring covers the same discipline in its full technical form.

Where the requirements meet the wiki

The recurring assignment requirements map directly onto the material:

• Grounded answers with citations and "I don't know" fallbacks: RAG
• Refusal logic and safety boundaries: Guardrails
• Strict output formats: the Structured Output page
• The eval harness that rubrics reward: Evaluation and Monitoring

A finished take-home also doubles as portfolio material, and the field guide recommends publishing it with context. It also recommends preparing by re-implementing published assignments and reading their READMEs before their code (take-home assignments). Portfolio Projects is the AISL view of what makes any project persuasive.

Where to go next

• The rest of the funnel: Interview Prep and AI Engineering Career
• Build the skills the rubrics pay for: AI Engineering Roadmap
• The measured interview process: Field Guide: Interviews