← Back to Projects
AgStream
Mentored by: Next Silicon
Real-time data streaming platform for agricultural analytics
Python
Apache Kafka
Apache Flink
InfluxDB
Grafana
Docker
Description
A scalable streaming data platform for processing agricultural sensor data in real-time. Enables continuous monitoring of crop conditions, weather patterns, and equipment status. Features include stream processing, time-series analysis, alerting, and integration with IoT devices for precision agriculture.
Team Members
Cohort: Embedded Systems Bootcamp 2025 (Embedded)
Responsibilities:
Designed and implemented a modular real-time inference pipeline on the Jetson platform using DeepStream and GStreamer for agricultural computer-vision tasks.
Integrated a PyTorch model, converted it to ONNX, and optimized it with TensorRT to run efficiently on the GPU with significantly reduced latency.
Developed custom preprocessing and post-processing logic, including background-removal capabilities and frame-filtering conditions for improved detection clarity.
Implemented frame-skipping and conditional processing strategies to maintain real-time performance under varying load and bandwidth conditions.
Built asynchronous processing flows enabling inference, storage, and streaming operations to run concurrently without blocking the main pipeline.
Added selective image-saving and integrated database logging mechanisms to store only relevant detection events while minimizing I/O overhead.
Developed performance-benchmarking tools to measure GPU utilization, inference latency, throughput, and dropped-frame rates across multiple configurations.
Set up a reproducible DeepStream development environment on the Jetson, including dockerized deployment, build scripts, and modular project structure.
Conducted accuracy validation by comparing PyTorch, ONNX, and DeepStream outputs, ensuring numerical correctness throughout optimization stages.
...and more contributions not listed here
Responsibilities:
Designed an AI inference pipeline for plant classification on edge device (NVIDIA Jetson).
Built a modular pipeline using DeepStream SDK and Python, including RTSP video decoding, CPU-based preprocessing (OpenCV), and GPU acceleration (C++ opencv.cuda) to avoid redundant memory transfers.
Integrated computer vision models (ResNet18 and MobileNet) converted PyTorch models to ONNX and optimized them with TensorRT for accelerated inference.
Inference results were stored in MongoDB via MQTT publishing for distributed deployment and offline analysis.
Development was done in a Linux environment using full Docker containers, with Git version control and code reviews by team peers and an experienced industry-grade team lead.
...and more contributions not listed here
Responsibilities:
Deep Learning Pipeline Development Developed an end-to-end inference pipeline using PyTorch → ONNX → TensorRT, integrated with NVIDIA DeepStream and GStreamer for real-time video processing and classification.
GPU Optimization & C++ Integration Implemented high-performance C++ modules with CUDA acceleration, achieving 31% performance improvement (120ms → 83ms for ResNet18). Created Python bindings using Pybind11 for seamless integration.
Computer Vision Algorithms Developed advanced image processing algorithms including ExG (Excess Green) background removal, Otsu thresholding, and morphological operations for improved plant detection accuracy.
Real-time Video Processing Built GStreamer-based video pipelines with RTSP streaming support, enabling real-time frame processing and intelligent frame skipping based on content analysis.
Edge Computing Deployment Optimized the system for NVIDIA Jetson platforms with ARM64 compilation, TensorRT optimization, and efficient memory management for field deployment.
Build System & DevOps Created advanced Makefile-based build system supporting CUDA 12.2, cross-compilation for ARM64, and automated d
Build System & DevOps Created advanced Makefile-based build system supporting CUDA 12.2, cross-compilation for ARM64, and automated dependency management for multiple target platforms.
Memory Management Optimization Implemented zero-copy memory management with direct EGL-CUDA interoperability, eliminating unnecessary data transfers between GPU and CPU memory spaces.
System Integration & Testing Integrated all components into a cohesive system with comprehensive error handling, logging, and performance monitoring for production deployment.
...and more contributions not listed here
Responsibilities:
Built a pipeline for plant classification, running at the edge on NVIDIA Jetson, and processing live video.
Integrated performance and accuracy optimizations, implemented in C++ to avoid GPU-to-CPU memory copies.
Evaluated and used computer vision models (ResNet18, MobileNet) for image analysis.
Set up an RTSP server to simulate live video during the development process.
Stored inference results in MongoDB via MQTT for later analysis.
Stored inference results in MongoDB via MQTT for later analysis.
...and more contributions not listed here