AI Engineer and Systems Developer specializing in clinical Generative AI, edge TinyML systems, and GPU-accelerated computing. Experienced in fine-tuning medical foundation models (MedGemma, MedSAM, BiomedCLIP), deploying quantized vision networks to microcontrollers, and building low-latency production pipelines using Python, C++, and CUDA.
An Energy-Efficient TinyML Framework for Real-Time On-Device Crop Disease Classification in Resource-Constrained Edge Environments
- Author: Azizbek Khadjibayev (Department of Infocommunication Engineering)
- Key Contributions:
- Engineered an ultra-compact depthwise separable Convolutional Neural Network (CNN) with 8,140 weights.
- Applied symmetric uniform 8-bit integer (INT8) Post-Training Quantization, fitting the complete model into a 34.95 KB Flash footprint (1.71% of available 2 MB Flash on ESP32-CAM) and 36.86 KB peak SRAM.
- Demonstrated 100.0% Top-1 accuracy and 1.0000 macro F1-score across foliar pathologies (Early Blight, Late Blight, Bacterial Spot, Healthy).
- Measured hardware execution latency of 2.9 ms and 0.76 mJ energy consumption per inference, enabling multi-year autonomous battery and solar operation.
- Code Repository: github.com/Zizx4uM/tinyml-crop-disease
- Clinical triage and medical question-answering assistant fine-tuned with Google's MedGemma.
- End-to-end symptom assessment, differential triage suggestions, and bilingual consultation flows with FastAPI backend and React frontend.
- Live Platform: https://shifoai.site
- Repository: github.com/Zizx4uM/shifoai-uzbek-mvp
- Enterprise credit surveillance platform monitoring 600+ corporate loan borrowers across industrial sectors.
- Combines continuous portfolio ingestion, statutory Central Bank regulatory trigger engines, and SHAP-explainable machine learning credit scoring models.
- Live Platform: https://www.nazoratai.uz
- Repository: github.com/Zizx4uM/NazoratAI
- On-device real-time agricultural diagnostics running quantized neural networks directly on microcontroller-class hardware.
- Complete hardware profiling, offline inference pipeline, and embedded sensor camera integration.
- Repository: github.com/Zizx4uM/tinyml-crop-disease
- Custom CUDA C++ kernels targeting NVIDIA Ada Lovelace architecture (RTX 4060).
- Parallel reductions, warp shuffle primitives, shared memory tiling for matrix multiplication, and memory coalescing optimization.
| Category | Badges |
|---|---|
| Languages |       |
| AI / Machine Learning |       |
| Medical AI Models |      |
| DevOps & Infrastructure |     |
