Structure

Structure is a local-first macOS desktop IDE for C++ technical interview training, with secondary Python support. It is designed to behave like a strict but helpful interviewer, tutor, debugger, and spaced-repetition coach rather than an autocomplete tool.

The primary goal is not to give answers. The primary goal is to help you become capable of solving DSA problems without assistance.

MVP Features

• Native Qt 6 / C++ desktop shell
• Four-panel IDE layout: problem, editor, test output, tutor/progress
• C++ and Python starter code support
• C++ sample and hidden test execution through clang++
• Python sample and hidden test execution through python3
• SQLite persistence for problems, attempts, settings, sessions, and review cards
• Built-in local problem pack with arrays, hashing, two pointers, sliding window, stack, binary search, graphs, trees, heaps, DP, intervals, and backtracking
• Mode selector for Learn, Interview, Debug, Speedrun, and Review
• Progressive hint ladder with dependency tracking
• Local OpenAI-compatible LLM chat endpoint integration
• One-command local Gemma 4 setup through Ollama
• Review card generation after each submitted attempt
• Basic progress dashboard and interview readiness estimate

macOS Dependencies

Install Qt 6 and CMake. With Homebrew:

brew install cmake qt

If you installed Qt with the Qt Online Installer, the CMake prefix is usually the kit folder, for example:

~/Qt/6.x.y/macos

You also need a C++ compiler. macOS Command Line Tools provide clang++:

xcode-select --install

SQLite is used through Qt SQL's SQLite driver.

Build

cmake -S . -B build -DCMAKE_PREFIX_PATH="$(brew --prefix qt)"
cmake --build build

With the Qt Online Installer path:

cmake -S . -B build -DCMAKE_PREFIX_PATH="$HOME/Qt/6.x.y/macos"
cmake --build build

Run the app:

open build/Structure.app

If you prefer launching the executable directly:

./build/Structure.app/Contents/MacOS/Structure

Local Gemma 4

Structure defaults to Gemma 4 through Ollama at:

http://127.0.0.1:11434/v1/chat/completions

The one-command setup installs Ollama if needed, starts the local server, chooses a performant Gemma 4 model for the machine, pulls it, warms it, keeps it resident, and updates any existing Structure SQLite settings.

scripts/setup-gemma4-local.sh

You can also run it through CMake:

cmake --build build --target setup-local-llm

Default model selection:

• speed profile: gemma4:e2b, the default for responsive tutoring
• balanced profile: also uses gemma4:e2b until larger models can consistently answer inside the 10-second budget
• quality profile: uses gemma4:26b on machines with at least 48 GiB RAM, otherwise gemma4:e2b

Override the model or profile:

STRUCTURE_GEMMA4_MODEL=gemma4:26b scripts/setup-gemma4-local.sh
STRUCTURE_MODEL_PROFILE=speed scripts/setup-gemma4-local.sh

Start the local server without pulling a model:

scripts/run-gemma4-local.sh

Check local model latency:

scripts/check-gemma4-local.sh

For fast first responses, Structure also warms the configured model once in the background when the app opens. When the configured endpoint is Ollama on port 11434, Structure automatically uses Ollama's native /api/chat path with think=false for tutor responses, plus keep_alive=1h, so the Gemma model returns concise final answers instead of spending the whole 10-second budget on reasoning text. Other endpoints continue to use the OpenAI-compatible /v1/chat/completions request shape. You can adjust timeout, max tokens, temperature, keep-alive, and startup warmup in Settings.

The scripts use Ollama on macOS because Ollama supports Apple M-series CPU/GPU execution and exposes an OpenAI-compatible /v1/chat/completions endpoint. For this app, that gives the simplest fast local path while preserving the existing local-first HTTP client.

Local LLM Server

Structure expects a local OpenAI-compatible chat endpoint.

Example endpoint:

http://127.0.0.1:11434/v1/chat/completions

You may run a local model using Ollama, llama.cpp, llama-cpp-python, MLX-backed tools, or another compatible server. Configure the endpoint and model name in Settings. The default model name is gemma4:e2b, the default endpoint is http://127.0.0.1:11434/v1/chat/completions, the default max response budget is 512 tokens, and the default timeout is 10 seconds.

Example llama.cpp-style server shape:

llama-server -m /path/to/model.gguf --port 8080

Ollama-compatible OpenAI endpoints often use:

http://localhost:11434/v1/chat/completions

How Problems Work

Built-in problems live in:

src/data/seed_problems.json

Each problem includes:

• title, difficulty, tags
• stdin/stdout problem statement
• examples and constraints
• C++ and Python starter code
• sample tests and hidden local tests
• expected time and space complexity
• pattern metadata and teaching notes

To add a problem, add another object to the problems array and rebuild the app. Existing local databases seed only when empty, so delete the local app database if you want to reload seed data during development.

Architecture

src/
  app/          QApplication setup and styling
  ui/           Qt Widgets IDE panels
  core/         Problem, Attempt, TestCase, ReviewCard, Session models
  execution/    C++ and Python subprocess runners
  ai/           Local LLM client, prompt builder, tutor engine, hint ladder
  storage/      SQLite database and repositories
  analysis/     MVP static-analysis hooks
  data/         Seed problems and prompt templates

Code execution is intentionally local. C++ attempts are written to a temporary directory, compiled with clang++, executed with a timeout, and compared against expected stdout.

Privacy

Structure is local-first. Your code, attempts, review cards, and interview practice history are stored locally on your machine. The app is designed to work with local LLMs and does not require cloud inference.

License

Structure is available under the MIT License. See LICENSE.

Non-Goals

Structure does not include cloud sync, accounts, social features, scraping, company-named prep packs, browser extensions, voice mode, or mobile support in the MVP.