Skip to content

Latest commit

 

History

History
103 lines (77 loc) · 3.71 KB

README.md

File metadata and controls

103 lines (77 loc) · 3.71 KB

Python → C++ Translation Agent

Translates Python to C++ and verifies correctness using a 3-agent loop, no compiler required.

How It Works

Python Code
    │
    ▼
┌─────────────┐
│  Translator │  Gemini — translates Python to optimized C++
└─────────────┘
    │
    ▼
┌──────────────────┐
│  Test Generator  │  GPT-5-Mini — reads the Python, produces expected behaviors
└──────────────────┘
    │
    ▼
┌───────────────┐
│   Evaluator   │  GPT-5-Mini — checks the C++ against expectations, flags issues
└───────────────┘
    │
    ├── PASS → done, save optimized.cpp
    └── FAIL/WARN → feed issues back to Translator (up to 3 attempts)

The translator and evaluator use different LLMs deliberately — if the same model translates and evaluates, it tends to confirm its own mistakes. Model choices are cost-driven; the architecture is model-agnostic so swapping in stronger models is a config change.

Quick Start

pip install -r requirements.txt
cp .env.example .env   # add your OpenAI + Gemini api keys
python main.py examples/pi_calculation.py --output output/pi.cpp

Or run the Gradio UI: python app.py

Two CPU-intensive examples are included: Leibniz π approximation (100M iterations) and maximum subarray sum with a custom RNG.

Background

This started as an exercise from Ed Donner's LLM Engineering course, where the task is to translate Python to C++ using an LLM and manually compile to verify. The two example scripts in examples/ are from the course material.

I wanted to automate verification without leaving Python; the agentic feedback loop, test generator, and evaluator are my additions to the original exercise.

Limitations

The evaluator does static analysis only. It reasons about the C++ but never executes it. The test generator predicts what the Python should output rather than running it. For the included examples (pure math, deterministic), this works well. For complex code, you'd want actual execution; e.g. OpenAI's Code Interpreter for Python ground truth, or a Docker container for C++ compilation.

Example Output

🔄 Step 1: Translating Python → C++...
🧪 Step 2: Generating test cases...
🔍 Step 3: Evaluating translation (attempt 1/3)...
✅ Translation passed evaluation on attempt 1.

📝 GENERATED C++ CODE
------------------------------------------------------------
#include <iostream>   // Required for input/output operations (std::cout, std::endl)
#include <chrono>     // Required for time measurement (std::chrono::high_resolution_clock)...

============================================================
  PYTHON → C++ TRANSLATION REPORT
============================================================
🧪 TEST CASES
------------------------------------------------------------

  Test 1: Run the program as written: compute result = calculate(100000000, 4, 1) * 4 and...


📊 EVALUATION
------------------------------------------------------------
  Verdict    : PASS
  Confidence : 90%
  Summary    : The C++ code implements the same algorithm as the Python code: 
               it iterates i=1..iterations, computes (1/(i*param1+param2) - 1/(i*param1-param2))
               accumulated into a double starting from 1.0, and multiplies 4.0 in main...


  Issues found:
    ⚠️  The translation uses the single-quote digit separator in the literal 100'000'000LL 
        which requires a C++14 (or newer) compiler. This is a compilation-standard...
    
============================================================