AgriEdge AI — Edge-Deployed Visual Risk Classification Platform
Designed a production-style edge AI inference system that performs low-latency visual risk classification at the edge and routes high-value cases into centralized data pipelines.
Highlights
- Built a clear separation between training, inference, and orchestration layers
- Implemented edge-optimized real-time and batch inference pipelines
- Designed a system that filters and routes data to reduce bandwidth and central compute cost
- Documented full system architecture, tradeoffs, and deployment model
Links
Impact
- Demonstrated a reusable edge-AI pattern applicable to insurance, manufacturing, and inspection systems
- Showed how tiered processing can improve latency, cost, and scalability
- Proved ability to design AI systems, not just train models
Context
In many real systems, sending all data to the cloud is expensive and slow. Most data is low-value; only a small fraction requires deeper processing.
This project demonstrates a tiered decision pipeline:
Classify at the edge → forward only high-risk or high-value cases → process centrally.
What I Built
A production-style edge AI system with:
- Separate layers for:
- Model training
- Edge inference
- Batch inference
- Orchestration and routing
- Real-time and batch image classification flows
- A pluggable decision-routing layer
- A full ingestion and preprocessing pipeline
The demo domain uses plant disease imagery, but the architecture is domain-agnostic and mirrors patterns used in insurance, manufacturing, and document processing systems.
Outcomes
- A reusable reference architecture for edge AI + data pipelines
- A working demonstration of low-latency, cost-aware compute routing
- A concrete example of AI system design beyond notebooks
Why This Matters
This project shows how I approach AI as infrastructure and systems engineering, not just model training:
Build decision pipelines, control cost and latency, and design for deployment from day one.