Open-Ended AI-Assisted Build: Admin Tool for User and Agent Conversation Logs

Quick Overview

A 70-minute open-ended full-stack build: given a folder of conversation logs between users and agents plus a manifest of user and agent attributes, and no requirements, decide what to build and build it with AI coding tools. It tests product judgment, choosing features an admin would need, and critically reviewing AI-generated code.

Open-Ended AI-Assisted Build: Admin Tool for User and Agent Conversation Logs

Company: Hotplate

Role: Software Engineer

Category: Software Engineering Fundamentals

Difficulty: easy

Interview Round: Onsite

In a 70-minute onsite round for a full-stack engineering role, you work on your own laptop and may use AI coding tools. The interviewer emails you a folder and says only: "There are no requirements. Build anything you see fit." Asked what the folder contains, the interviewer tells you to figure it out, perhaps by asking an AI assistant. No stack, product or deliverable is specified, and later requests for requirements get the same answer: build whatever you want. The folder holds a small dataset: - **Ten conversation logs.** Each log is a JSON array containing a conversation between users and agents. - **`manifest.json`.** It describes the participants: for each user, a user id (`userId`), an avatar image URL and a conversation id; for each agent, attributes such as the agent's role and its display color. Decide what to build, build it within the round, and explain your choices as you go. If you propose a direction, for example an internal admin tool for inspecting the logs, expect the interviewer to accept it and then probe it: why that form, and what features its users would want. ### Constraints and Clarifications - 70 minutes in total, on your own machine; AI coding assistants are allowed. - Stack, architecture and scope are yours to choose. - Accessibility and mobile responsiveness are not required. - Integration tests are not required. - The report does not give the field names inside the logs; inspect the files to find them. ### Clarifying Questions - Is a backend in scope, with services or endpoints that process the logs, or is a front end over the raw files enough? - Who would use what you build: the team that builds the agents, support or operations staff, or someone else? - Are the agents automated agents or human staff, and can several agents take part in one conversation? - How do the logs link to the manifest: does each log correspond to exactly one conversation id, and can a user appear in more than one conversation? - Is a working demo expected at the end of the 70 minutes, or is a partial build with a clear plan acceptable? ### Part 1 — Understand the data and choose what to build Work out what the folder contains and how the logs and the manifest relate, then propose what you will build, for whom, and why. Be ready to defend the shape of the main view: if you propose a table of conversations, explain why a table, and what an admin would do with it. ```hint Read it yourself Open `manifest.json` and one complete log alongside the assistant's summary, and find the identifiers that connect a message to a user or an agent before you decide anything. ``` #### What This Part Should Cover - How the data was inspected and verified, not only summarized by the AI - The join between the logs and the manifest, and any gaps or inconsistencies in it - A concrete user and job for the tool, stated as a short proposal - A justification for the main view's form, given ten conversations now and possibly many more later ### Part 2 — Choose features that matter The interviewer asks what features an admin would want. Propose the features, rank them, and decide which ones you will build in the remaining time. Rendering the conversations is the baseline; explain what you would add beyond it that makes the tool useful for this particular data. ```hint Start from the admin's questions List the questions someone responsible for these agents would bring to the tool, then check which of them the logs and the manifest can answer that a plain list of rows cannot. ``` #### What This Part Should Cover - Features tied to questions this data can actually answer, including information derived from it - How the manifest's user and agent attributes are used in the views - A prioritized plan that fits the 70 minutes, and what was deliberately left out - Data problems the tool itself should surface ### Part 3 — Build with AI and own the result Build the tool with your AI assistant. Explain how you direct it, how you review what it produces, and what you do with generated code you do not want. Integration tests are not required; decide what checking you will do instead. ```hint Review before you accept Decide what you will check in each generated change before you accept it, and what should happen to generated components you do not need. ``` #### What This Part Should Cover - Breaking the build into small, specified requests instead of one broad prompt - Reading and running every generated change, and being able to explain the code when asked - Removing unneeded generated code rather than leaving it in the codebase - Lightweight verification of the data handling, given that integration tests were waived ### What a Strong Answer Covers - Takes ownership of the ambiguous brief: proposes a direction early, confirms it, and commits to it - Understands and verifies the data before building on it - Builds features that are meaningful for conversations between users and agents, not generic dashboard scaffolding - Scrutinizes AI-generated code and can explain every part of the final result - Uses the full-stack scope: decides deliberately where processing belongs, front end or back end - Finishes with a working demo and a clear account of trade-offs and next steps ### Follow-up Questions - The ten logs become millions of conversations arriving daily from production. What changes in how you load, search and page through them? - Pick one component the AI generated and walk through it line by line. What would you change before merging it? - An admin wants to mark a conversation as problematic and share it with a teammate. What does that need on the backend and in the UI? - Message text may contain markup or links typed by users. How do you render it safely?

Overview: A 70-minute open-ended full-stack build: given a folder of conversation logs between users and agents plus a manifest of user and agent attributes, and no requirements, decide what to build and build it with AI coding tools. It tests product judgment, choosing features an admin would need, and critically reviewing AI-generated code.

Read the full Hotplate Software Engineer interview experience this question came from

Hotplate logo
Hotplate
Nov 16, 2026
easySoftware EngineerOnsiteSoftware Engineering Fundamentals
0
0

In a 70-minute onsite round for a full-stack engineering role, you work on your own laptop and may use AI coding tools. The interviewer emails you a folder and says only: "There are no requirements. Build anything you see fit." Asked what the folder contains, the interviewer tells you to figure it out, perhaps by asking an AI assistant. No stack, product or deliverable is specified, and later requests for requirements get the same answer: build whatever you want.

The folder holds a small dataset:

  • Ten conversation logs. Each log is a JSON array containing a conversation between users and agents.
  • manifest.json. It describes the participants: for each user, a user id ( userId ), an avatar image URL and a conversation id; for each agent, attributes such as the agent's role and its display color.

Decide what to build, build it within the round, and explain your choices as you go. If you propose a direction, for example an internal admin tool for inspecting the logs, expect the interviewer to accept it and then probe it: why that form, and what features its users would want.

Constraints and Clarifications

  • 70 minutes in total, on your own machine; AI coding assistants are allowed.
  • Stack, architecture and scope are yours to choose.
  • Accessibility and mobile responsiveness are not required.
  • Integration tests are not required.
  • The report does not give the field names inside the logs; inspect the files to find them.

Clarifying Questions Guidance

  • Is a backend in scope, with services or endpoints that process the logs, or is a front end over the raw files enough?
  • Who would use what you build: the team that builds the agents, support or operations staff, or someone else?
  • Are the agents automated agents or human staff, and can several agents take part in one conversation?
  • How do the logs link to the manifest: does each log correspond to exactly one conversation id, and can a user appear in more than one conversation?
  • Is a working demo expected at the end of the 70 minutes, or is a partial build with a clear plan acceptable?

Part 1 — Understand the data and choose what to build

Work out what the folder contains and how the logs and the manifest relate, then propose what you will build, for whom, and why. Be ready to defend the shape of the main view: if you propose a table of conversations, explain why a table, and what an admin would do with it.

What This Part Should Cover Guidance

  • How the data was inspected and verified, not only summarized by the AI
  • The join between the logs and the manifest, and any gaps or inconsistencies in it
  • A concrete user and job for the tool, stated as a short proposal
  • A justification for the main view's form, given ten conversations now and possibly many more later

Part 2 — Choose features that matter

The interviewer asks what features an admin would want. Propose the features, rank them, and decide which ones you will build in the remaining time. Rendering the conversations is the baseline; explain what you would add beyond it that makes the tool useful for this particular data.

What This Part Should Cover Guidance

  • Features tied to questions this data can actually answer, including information derived from it
  • How the manifest's user and agent attributes are used in the views
  • A prioritized plan that fits the 70 minutes, and what was deliberately left out
  • Data problems the tool itself should surface

Part 3 — Build with AI and own the result

Build the tool with your AI assistant. Explain how you direct it, how you review what it produces, and what you do with generated code you do not want. Integration tests are not required; decide what checking you will do instead.

What This Part Should Cover Guidance

  • Breaking the build into small, specified requests instead of one broad prompt
  • Reading and running every generated change, and being able to explain the code when asked
  • Removing unneeded generated code rather than leaving it in the codebase
  • Lightweight verification of the data handling, given that integration tests were waived

What a Strong Answer Covers Guidance

  • Takes ownership of the ambiguous brief: proposes a direction early, confirms it, and commits to it
  • Understands and verifies the data before building on it
  • Builds features that are meaningful for conversations between users and agents, not generic dashboard scaffolding
  • Scrutinizes AI-generated code and can explain every part of the final result
  • Uses the full-stack scope: decides deliberately where processing belongs, front end or back end
  • Finishes with a working demo and a clear account of trade-offs and next steps

Follow-up Questions Guidance

  • The ten logs become millions of conversations arriving daily from production. What changes in how you load, search and page through them?
  • Pick one component the AI generated and walk through it line by line. What would you change before merging it?
  • An admin wants to mark a conversation as problematic and share it with a teammate. What does that need on the backend and in the UI?
  • Message text may contain markup or links typed by users. How do you render it safely?