Modular ML Pipeline for Multi-Modal Classification (AI-Generated Content Detection)

Modular platform for ML research across multiple domains: food prediction (recipe difficulty, meal type, nutrients, time) and AI-generated content detection (deepfakes, image authenticity). Integrates advanced augmentation, unified pipelines, automated benchmarking, third-party research code (FatFormer, DiffusionFake, TabR, GRANDE, TabM), two standalone apps (a conversational nutritionist agent and a recipe analysis lab), and a research framework for AI-content detection and adversarial evasion across text and image.

Project Structure

• ml_pipeline/: Core ML experimentation suite (augmentation, pipelines, benchmarking, reproducibility, results). See ml_pipeline/README.md
• nut_agent/: Two Streamlit apps: NutriCoach (LLM-powered nutritionist agent) and Recipe Lab (ML-powered recipe analyzer). See nut_agent/README.md
• ai_content_detector/: Research framework for AI-content detection and adversarial evasion across text and image. Provides an ensemble of zero-shot and supervised detectors, RL-based evasion trainers (GRPO, MultiSPIN, DDPO), a multi-round attacker/defender arms race, and a Streamlit scoring app. Reuses ml_pipeline/ checkpoints and wrappers. See ai_content_detector/README.md

Quickstart

1. Install dependencies: uv pip install -r requirements.txt for the standard environment, or uv pip install -r requirements_gpu.in on the CUDA 12.4 GPU VM used for notebook training.
2. Interactive Benchmark UI: cd ml_pipeline && streamlit run benchmark_app.py (with OpenMP configured)
3. Launch NutriCoach: cd nut_agent && streamlit run nutricoach/app.py
4. Launch Recipe Lab: cd nut_agent && streamlit run recipe_lab/app.py
5. Launch AI Content Detector: streamlit run ai_content_detector/app.py

Benchmark App (Streamlit)

Interactive UI for model benchmarking with registry-based configuration:

• Launch: cd ml_pipeline && streamlit run benchmark_app.py
• Features: Upload CSV datasets, select models/augmentations from registries, configure training parameters, visualize results
• Location: ml_pipeline/benchmark_app.py

macOS Setup (for XGBoost): Install OpenMP via MacPorts (sudo port install libomp) and set DYLD_LIBRARY_PATH permanently in your shell config.

Example (bash):

echo 'export DYLD_LIBRARY_PATH="/opt/local/lib/libomp:$DYLD_LIBRARY_PATH"' >> ~/.bash_profile
source ~/.bash_profile

Testing

# ml_pipeline (73 tests: scoring, wrappers, pipelines, checkpoints, benchmarks)
cd ml_pipeline && python -m pytest tests/ -v

# nut_agent (86 tests: utils, auth, memory, predictor, intent, agent)
cd nut_agent && python -m pytest tests/ -v

# ai_content_detector (146 tests: detectors, rewards, GRPO/MultiSPIN math,
# arms-race equilibrium, WaRPAD, ensemble — pure-Python, CPU-only)
PYTHONPATH=. python -m pytest ai_content_detector/tests/ -v

Notes

• User data and chat logs: nut_agent/secrets/ (excluded from version control)
• ML models and results: ml_pipeline/results/
• Registry-based design for easy addition of models, augmentations, and metrics

For full details, consult the README in each subfolder.