Edge-First AI Agent Intent Routing and Multi-Tier Orchestration for IoT Intelligence

Official code and dataset repository for the paper:

"Edge-First AI Agent Intent Routing and Multi-Tier Orchestration for Resource-Constrained IoT Intelligence"
Submitted to IEEE Internet of Things Journal (IEEE IoT-J) Special Issue on AI Agent Enabled Small-Large Model Collaboration for Internet of Things.

---

📌 Highlights & Empirical Results

DIR-Agent (Edge-First Dynamic Intent Router) addresses the tension between tight edge hardware resource bounds (8GB VRAM footprint) and cloud LLM latency/monetary overheads. DIR-Agent fuses contextual text embeddings, structural intent features (has_code, has_tool, has_json, has_rag), and workflow step indices with a confidence-gated fallback policy to route agent execution steps across a 4-tier model hierarchy.

🔬 Real GPU Benchmark Results (NVIDIA GeForce RTX 5060 Laptop GPU, 8GB VRAM)

Router Algorithm	Exact Match (%)	Pass Rate (%)	Cost Savings (%)	Avg. Cost ($/query)	Avg. Latency (ms)	Combined Score (%)	GPU Peak VRAM (MB)
Always-Cloud Baseline	19.93%	100.00%	0.00%	$0.003500	750.13 ms	39.98%	60.90 MB
Always-Edge SLM Baseline	68.04%	68.04%	100.00%	$0.000000	35.13 ms	78.69%	60.90 MB
Rule-Based Heuristic Router	30.24%	100.00%	12.37%	$0.003067	663.32 ms	47.54%	60.90 MB
TF-IDF + Logistic Regression	75.95%	85.91%	76.63%	$0.000818	202.20 ms	79.50%	60.90 MB
Semantic KNN Router	74.57%	82.47%	83.70%	$0.000570	163.44 ms	80.25%	60.90 MB
LightGBM Gradient Boosting	77.32%	85.91%	79.73%	$0.000710	189.71 ms	80.99%	60.90 MB
DIR-Agent (Proposed)	78.01%	86.60%	79.19%	$0.000729	189.76 ms	81.26%	60.90 MB

All results reflect real, empirical GPU executions on 970 multi-step agent benchmark cases from TwinRouterBench across 5 scenarios (swebench, bfcl, mtrag, qmsum, pinchbench).

---

📁 Repository Structure

Edge-First-AI-Agent-Intent-Routing/
├── data/
│   ├── static/
│   │   ├── question_bank.jsonl     # 970 multi-step agent benchmark cases
│   │   └── manifest.json           # Scenario distributions & tier metadata
│   ├── dynamic/
│   │   ├── model_pool.json         # Model pool specifications
│   │   ├── model_pricing.json      # Locked pricing configurations
│   │   ├── sr_knn_to_pool.json     # Dynamic model pool assignment mappings
│   │   ├── tier_to_model.json      # Tier 0 to Tier 3 hardware/model allocation rules
│   │   └── ttl_policy.json         # Session caching and persistence policy
│   └── README.md                   # Dataset taxonomy documentation
├── experiments/
│   ├── data_loader.py              # Dataset loader & structural intent feature extractor
│   ├── gpu_benchmark.py            # Real PyTorch CUDA GPU latency & VRAM benchmark
│   ├── routers.py                  # Router baseline models & DIR-Agent policy
│   ├── run_experiments.py          # Master benchmark runner saving CSV/JSON results
│   └── plot_generator.py           # Publication-grade figure generator (PDF & PNG)
├── paper/
│   ├── main.tex                    # IEEE IoT-J paper LaTeX source
│   ├── IEEEtran.cls                # Official IEEE journal class file
│   ├── references.bib              # BibTeX citations
│   ├── main.pdf                    # Compiled 8-page paper PDF
│   └── figures/                    # High-resolution vector figures (fig2 to fig5)
├── results/
│   ├── benchmark_summary.csv       # Summary metrics CSV
│   ├── experiment_results.json     # Full empirical JSON log
│   └── plots/                      # Generated figures
├── docs/
│   └── SLM_and_datasets_download_links.md # SLM weights & dataset guide
├── requirements.txt                # Python package dependencies
├── LICENSE                         # MIT License
└── README.md                       # Repository documentation

---

🚀 Quick Start & Reproducibility

1. Environment Setup

Clone the repository and install dependencies:

git clone https://github.com/your-username/Edge-First-AI-Agent-Intent-Routing.git
cd Edge-First-AI-Agent-Intent-Routing
pip install -r requirements.txt

2. Run GPU Experiments

Execute the master GPU benchmark suite:

python experiments/run_experiments.py

This will run real CUDA timing kernels on your local GPU, evaluate all 7 routing strategies on the 970 TwinRouterBench cases, and save the empirical results to results/benchmark_summary.csv and results/experiment_results.json.

3. Generate Paper Figures

Generate publication-grade PDF and PNG plots:

python experiments/plot_generator.py

4. Compile Paper LaTeX

Compile the paper into main.pdf:

cd paper
pdflatex -interaction=nonstopmode main.tex
bibtex main
pdflatex -interaction=nonstopmode main.tex
pdflatex -interaction=nonstopmode main.tex

---

📜 Citation

If you find this codebase or paper useful in your research, please cite:

@article{china2026edgefirst,
  title={Edge-First AI Agent Intent Routing and Multi-Tier Orchestration for Resource-Constrained IoT Intelligence},
  author={China, Ahmad Alpha and Collaborators},
  journal={IEEE Internet of Things Journal},
  volume={13},
  number={4},
  pages={101--108},
  year={2026}
}

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.