Back to interview questions

Interview Questions

Home Assignments

Take-home assignments, paid work trials, and asynchronous assessments for AI/ML engineering roles. Based on 1,765 job descriptions, 100+ GitHub repos of actual candidate submissions, and practitioner reports.

Take-home assignments, paid work trials, and asynchronous assessments for AI/ML engineering roles. Based on 1,765 job descriptions, 100+ GitHub repos of actual candidate submissions, and practitioner reports.

Of the 51 companies with disclosed interview processes, 17 (33%) include a take-home or asynchronous assignment. An additional 5 companies use paid work trials instead. Analysis of 100+ GitHub repos (Q4 2025 / Q1 2026) shows what companies actually ask:

  • RAG systems (40%+) - document upload, vector databases, citation support
  • Agentic systems (30%+) - tool-calling, multi-step reasoning, multi-agent orchestration
  • Conversational AI (20%+) - chatbots, live chat agents, voice assistants
  • Document processing (15%) - PDF parsing, data extraction, marksheet extraction
  • LLM-as-judge evaluation (10%+) - build a system then evaluate it with another LLM

Format

Asynchronous assignments completed on your own time, typically with a deadline of 2-7 days. You submit code, a writeup, or a working prototype, then discuss your solution in a follow-up interview.

Typical assignment formats:

  • Take-home coding project (most common) - build something from scratch, typically 2-4 hours of actual work. "Not code golf" - companies emphasize decision-making and clarity over cleverness
  • Defence round - present and defend your solution in a 45-90 minute walkthrough interview

Only 1 company explicitly allows AI tools in take-homes ("AI tools are explicitly allowed and encouraged"). No company explicitly bans AI for take-home assignments (bans only apply to live interviews). Most don't mention AI tool policy at all.

Assignment Examples

Actual take-home assignments from candidate reports, job listings, and 100+ GitHub repos of real submissions.

How to Prepare

The best submissions share these traits:

  • Ask clarifying questions before starting. Double your time estimate.
  • Start with evaluation - build an eval harness before writing the main logic. YC startups report this as the top signal: "Red flag if candidate doesn't start with evals"
  • Document design decisions and trade-offs - why you chose this approach over alternatives
  • Include a Loom/video walkthrough of your submission
  • Make it configurable - one engineer built a PDF summarizer CLI with a config file for different models and chunking strategies, got two competing offers within 72 hours
  • Test edge cases - even if testing isn't explicitly required, include basic tests
  • Show production awareness - error handling, monitoring hooks, cost estimates
  • Connect technical metrics to business outcomes
  • Prepare for the defence round - practice explaining your architecture decisions, trade-offs, and what you'd do differently with more time

Common mistakes:

  • Not asking clarifying questions before starting
  • Putting too little effort in - "most engineers put way too little effort into take-homes"
  • Rushing without documenting design decisions and trade-offs
  • Over-engineering beyond what was asked without justification
  • Ignoring evaluation and testing of AI outputs - this is the single biggest differentiator
  • Not preparing for the walkthrough/defence round - the follow-up interview is often more important than the code itself

June 2026 Additions

New assignments discovered in a GitHub search pass covering Q2 2026. These extend the patterns above and introduce several new assignment archetypes.

RAG and Document Q&A

The most common assignment type (40%+ of GitHub submissions).

  1. 1 Build a RAG chatbot that ingests PDFs/documents, creates embeddings in a vector DB, and answers questions with citations. Must respond "I don't have that information" when answer is unavailable. Answers must come strictly from retrieved context (10+ candidate submissions across 5+ companies).
  2. 2 Build a policy document RAG assistant with mandatory source citations for every answer. Return safe fallbacks for out-of-scope questions. Comes with a 7-question evaluation set across 3 categories (answerable, partially answerable, unanswerable).
  3. 3 Build a document Q&A system with citation tracking that handles multi-hop questions (questions requiring information from multiple documents or sections to answer).
  4. 4 Build a live chat agent grounded in FAQ knowledge base. Model must answer only from known FAQ data.
  5. 5 Design a customer support chatbot using RAG with open-source models. Requirements: 100+ concurrent users, <2 second latency, grounded in company docs, analytics tracking.
  6. 6 Build a CLI tool for summarizing long PDFs with configurable models and chunking strategies.
  7. 7 Refactor an existing messy RAG application into a clean architecture. Preserve all external behaviors (exact API endpoints), eliminate global mutable state, ensure testability without requiring running services (5+ candidate submissions for one company).
  8. 8 Build an agentic RAG system for government documents. 100% open-source required (Ollama + CrewAI + pgvector). Must integrate with OpenWebUI. Evaluated using RAGAS metrics (faithfulness, answer relevancy, context precision, context recall).

Agents and Tool-Calling

Second most common (30%+).

  1. 1 Build an assistant agent handling database queries, document search, and bash commands. Bash commands require explicit user approval.
  2. 2 Build an AI agent that transforms Monday.com project management data into conversational business insights using dual-LLM architecture.
  3. 3 Build an AI agent demonstrating natural interaction, agentic behavior, and clear reasoning steps.
  4. 4 Build a customer support agent.
  5. 5 Build an autonomous agent using an open-source LLM with observability/eval layer.
  6. 6 Build an agent that reads customer CSV data and generates personalized email campaigns with evaluation metrics.
  7. 7 Build a code review agent that analyzes Python files and provides actionable feedback.
  8. 8 Build a Singapore public transport query agent that fetches live data from 7 LTA APIs about buses, trains, traffic, and station conditions.
  9. 9 Build a sales insights agent that answers questions about subscription/revenue data. Must detect and refuse PII requests (emails, phone numbers, credit card tokens). No raw rows passed to the LLM - aggregates only. Evaluated on 3 dimensions: accuracy, safety/refusal correctness, reasoning quality.
  10. 10 Build an evaluation tool for LLM hallucination detection.

Multi-Agent Systems

  1. 1 Build a multi-agent content generation system: 5 core agents (research, writing, editing, SEO, publishing). Takes product JSON input, generates FAQ document, product page, and comparison page. All outputs must follow strict JSON formats. LangChain + Groq.
  2. 2 Implement a minimal workflow engine with graph-based nodes, state management, branching/looping, and tool-based logic. Max 50 steps. Built-in infinite-cycle protection required. Unit tests mandatory (6+ candidate submissions).
  3. 3 Build a 5-agent CBT therapy system: agents autonomously design, critique, and refine therapy exercises. Human-in-the-loop approval required before finalization.
  4. 4 Build a 4-stage bedtime story pipeline: Spec Builder, Storyteller, LLM Judge, Rewriter. Must use gpt-3.5-turbo. Up to 2 revision cycles. LLM judge evaluates stories against the spec.

Document Extraction and Processing

  1. 1 Build a marksheet extraction API: parse complex table layouts and handwriting from academic marksheets into structured JSON.
  2. 2 Build a physician notetaker: transform physician-patient conversations into structured clinical documentation..
  3. 3 Build a legal document analysis tool for contracts: extract key information, identify risks (auto-renewal traps, liability, IP ownership, non-competes), generate structured summaries.
  4. 4 Build a CBT assistant combining RAG with safety mechanisms. Crisis detection mandatory. PII redaction and pseudonymization required. No secrets in logs. Educational only, not clinical advice.
  5. 5 Take a blood test report as PDF, understand medical issues, generate suggestions by fetching content from online blog articles with source links.
  6. 6 Build a question deduplication and clustering pipeline: exact dedup, semantic dedup, LLM-based cluster discovery, classification. Output evaluated with ARI, NMI, homogeneity, completeness metrics.
  7. 7 Build a data pipeline that processes 1,000 messy products from 4 vendors, normalizes them into a unified schema, fetches supplementary data via rate-limited async API calls (vendor-specific token-bucket rate limits), enriches products through AI-powered duplicate detection. Must support both CLI and API access.
  8. 8 Build a transaction-to-user matching system: identify users whose names appear in transaction descriptions, find similar transactions via text matching, propose improvements (semantic embeddings, database integration). Spring Boot + Java.
  9. 9 Build real-time earnings call transcription and insight streaming: streaming audio-to-text via Whisper, real-time extraction of financial signals (revenue, guidance, risks, outlook), SSE output.

Full-Stack AI Applications

  1. 1 Build a Telegram bot for investment coaching with safety filtering. Educational content only - no personalized financial advice.
  2. 2 Build a multi-agent D&D dungeon simulation: Game Master agent + Player agents. Must address at least 3 of 6 challenges (long campaigns, secrets, rulings, self-aware dungeon, living world, ambiguity). LangGraph required.
  3. 3 Build a memory extraction and personality transformation system: extract structured long-term memory from chat history as JSON, transform responses based on personas (calm mentor, witty friend, therapist). Open-source LLMs only, no proprietary APIs.
  4. 4 Build an AI judge for a Rock-Paper-Scissors variant: classify player inputs as VALID/INVALID/UNCLEAR, handle typos and edge cases, tool-based state management workflow (2 submissions).
  5. 5 Build a web app that converts markdown to slide deck presentations by splitting content into logical sections based on a target slide count. Document size limited to 150K tokens, single API call. Next.js + OpenAI.
  6. 6 Build an LLM processing pipeline with intelligent routing, multi-level caching (exact + semantic), provider health monitoring with failover, and distributed tracing. Targets: 100+ req/s, p95 latency under 2s, >40% cache hit rate.
  7. 7 Build an NPC system for a job simulation platform: three AI co-workers with distinct personalities, a "Director Agent" that detects conversation loops via semantic similarity (0.85 threshold), RAG-based knowledge retrieval. FastAPI + Claude API + FAISS.
  8. 8 Build an LLM-based rating prediction and prompt evaluation system with user and admin dashboards. Node.js + MongoDB + Google Generative AI.
  9. 9 Build an AI-first CRM module: React/Redux frontend, FastAPI backend, LangGraph with 5+ tools. Deliverable: GitHub repo + 10-15 minute demo video.

Evaluation Criteria Found in Assignments

Many assignments include explicit evaluation criteria. Patterns across repos

  1. 1 Functional correctness - does the system work end-to-end, handle edge cases, produce correct outputs
  2. 2 Code quality and architecture - modular design, clean code, extensibility, proper error handling
  3. 3 Evaluation methodology - whether candidates build eval harnesses, define metrics, measure quality systematically
  4. 4 Production readiness - scalability considerations, caching, monitoring, cost optimization, security (PII handling, input sanitization, rate limiting)
  5. 5 Performance targets - response time (e.g., <2s p95 latency, ~0.5s per response), throughput (100+ req/s), cache hit rates (>40%), cost reduction (>30%)
  6. 6 Testing - unit tests (sometimes mandatory), test coverage targets (e.g., 80%), edge case handling
  7. 7 Documentation - README quality, design decision explanations, trade-off analysis
  8. 8 Weighted rubrics - some assignments provide explicit scoring: e.g., 30% functionality, 30% challenge completion, 25% context engineering, 15% code quality
  9. 9 Quantitative metrics - ARI/NMI for clustering , RAGAS metrics for RAG (faithfulness, answer relevancy, context precision) , confidence scores for extraction

RAG and Document Q&A (New Companies)

  1. 1 Quorium: Build a RAG Q&A chatbot for an AI Engineer Trainee role. Standard RAG pipeline with document ingestion and question answering.
  2. 2 ITJ: RAG-based document QA system as a take-home challenge.
  3. 3 NTT DATA: RAG system over sustainability reports.
  4. 4 NeoStats: Chatbot assignment for an AI Engineer case study.
  5. 5 Trinamix: Supply chain RAG chatbot over a 2,000-PO supplier register and governance policy. Built with Flowise, Pinecone, and GPT.
  6. 6 GoTyme Bank: Full-stack document extraction system combining OCR and LLM, with a React frontend.

Agents and Multi-Agent Systems (New)

  1. 1 Go Fig AI: Build an inbox-triage agent skill with a human-in-the-loop approval gate. 2-hour cap.
  2. 2 Yuno: Multi-agent orchestration platform with Ollama, Streamlit, and Telegram.
  3. 3 RefundPilot: Containerized internal support workspace that evaluates e-commerce refund requests, applies refund policy, resists prompt-injection attempts, and records structured decision logs.
  4. 4 AgentCollect: Full-Stack AI Engineer (AI-Native) hiring challenge.
  5. 5 Neon Health: AI Agent with OpenAI integration.
  6. 6 Future Research: LangGraph multi-agent fitness coach with hub routing to sub-agents (coach, workout-generator, workout-logger) via Claude structured output.
  7. 7 KarthikTools: Full-stack web app where a backend agent executes predefined tools with a clear execution trace.

LLM Applications and Infrastructure

  1. 1 GenAI Labs: Production-ready LLM-driven SQL analytics pipeline with token counting, SQL validation, observability, and benchmarking. Multiple submissions.
  2. 2 VantageScore: GenAI-powered credit risk scoring with ML ensemble, explainability, and LLM enrichment.
  3. 3 AEGIS: AEO content scoring, LLM query fan-out, and embedding-based semantic gap analysis.
  4. 4 SHL: GenAI assessment recommendation system.
  5. 5 Camplight: LLM interview task.
  6. 6 VijaySaravanaPandi: NL-to-app "compiler" — turns a natural-language app description into a working application via structured, validated intermediate representations (not just prompt to code).
  7. 7 Jpower3145: Local LLM pre-interview task.
  8. 8 Cerebras: AI Engineer Model Quality and Performance hiring challenge (perf UI, eval pruning).
  9. 9 nickusevich: Novelty detection for football news using hybrid retrieval (pgvector + tsvector + RRF), LLM reranking, and LLM-based publish/skip/review decisions.

Embeddable and Conversational AI

  1. 1 EloquentAI: Build an embeddable AI agent chat widget.
  2. 2 Fleetio: Hybrid deterministic + LLM weekly fleet digest.
  3. 3 Zap: AI-powered client onboarding automation.
  4. 4 Adobe (FDE): GenAI-powered creative automation pipeline for social ad campaigns.

Official Company Challenges

Official repos published by companies as part of their hiring process

  1. 1 [ML6 (laine)](https://github.com/ml6team/laine-engineer-coding-challenge) - Coding challenge for evaluating AI Engineer candidates
  2. 2 [Jaseci Labs](https://github.com/jaseci-labs/take-home-ai-engineer) - Take-home for AI Software Engineer candidates
  3. 3 [Jitera](https://github.com/Jitera-Interviews/genai-takehome) - Take-home task for GenAI roles
  4. 4 [AuxoAI](https://github.com/AuxoAI-Hiring/ai-engineer-assignment) - AI Engineering Take-Home
  5. 5 [Coginis Research](https://github.com/CoginisResearch/ai-engineer-challenge) - AI Product Engineer Hiring Challenge
  6. 6 [Future Research](https://github.com/future-research/candidate-assessment) - AI Engineer take-home assessment + exercise dataset
  7. 7 [Go Fig AI](https://github.com/go-fig-ai/take-home-inbox-triage) - Inbox-triage agent with human-in-the-loop
  8. 8 [Cerebras](https://github.com/danielkim-cerebras/ai-model-quality-challenge) - Model Quality and Performance challenge
  9. 9 [AI:AT](https://github.com/AIAT-AIandBusinessgrowth/standort-agent-challenge-public) - Multi-criteria location evaluation agent
  10. 10 [Bloom (radialreview)](https://github.com/radialreview/bloom-coffee-ai) - Pre-built coffee ordering app candidates extend with an AI order taker

Source: AI Engineering Field Guide (interview/questions/06-home-assignments.md)