🌴 Palm Oil FFB Management System (YOLO26)

⚠️ PROTOTYPE / PROOF OF CONCEPT (PoC) > This system is developed strictly for demonstration purposes to potential clients and is not intended for production use in its current state.

A high-fidelity AI prototype for detecting the ripeness of Palm Oil Fresh Fruit Bunches (FFB) according to MPOB standards. Built on a custom-trained YOLO26 model (utilizing the latest high-performance architecture via ultralytics) with a triple-engine inference backend and a specialized 5-tab operations dashboard.

LINK for GOOGLE DOC: https://docs.google.com/document/d/1bbdGuLCGHdnHj_iCQIreh9DlYwu12PDUEPpA2eCsR-M/edit?usp=sharing

🚀 Project Overview

Component	Technology	Purpose
Primary Engine	YOLO26 (Latest SOTA Backbone)	High-accuracy FFB Ripeness Detection.
ONNX Runtime	`onnxruntime` + `best.onnx`	Ultra-fast edge inference simulation (~39ms).
PyTorch Runtime	`ultralytics` + `best.pt`	High-resolution auditing and verification.
Benchmark Engine	YOLOv8-Sawit (`sawit_tbs.pt`)	Competitive baseline comparison.
Inference Server	FastAPI (Python)	REST API for multi-client orchestration.
Local Audit Log	SQLite (`palm_history.db`)	Primary offline-first data persistence.
Operations UI	Streamlit (`demo_app.py`)	5-tab industrial management dashboard.

🛠️ Triple-Engine Inference Logic

The system demonstrates three distinct AI pathways to showcase deployment flexibility:

YOLO26 (ONNX): Optimized for zero-latency scenarios. It utilizes normalized coordinates ($0.0$ to $1.0$) mapped back to original resolution via a custom scaling handler.
YOLO26 (PyTorch): The native "Ground Truth" engine. It provides absolute pixel coordinates and serves as the benchmark for high-resolution audits.
YOLOv8-Sawit (Benchmark): A third-party model implementation. It allows clients to see how the custom YOLO26 architecture performs against standard industry baselines.

🖥️ Demonstration Features (Streamlit)

Tab 1: Single Analysis: Drag-and-drop detection with interactive Plotly overlays and a "Manager's Dashboard" for OER (Oil Extraction Rate) yield-loss insights.
Tab 2: Batch Processing: Simulates a full harvest day by processing multiple images under a specific metadata context (Estate, Block, Harvester) and generating an Executive PDF Report.
Tab 3: Similarity Search (KIV): Keep In View. Currently staged for future integration. Designed for visual and natural language search via Vertex AI.
Tab 4: History Vault: A local audit log showing every analysis performed. It displays archived images alongside their raw mathematical tensors.
Tab 5: Batch Reviewer: An auditor's tool to browse manifest.json data contracts and verify resolution-agnostic norm_box coordinates.

🚧 Status & Future Roadmap (KIV)

The following features are currently KIV (Keep In View) and are disabled in the current prototype due to cloud service constraints:

Cloud Archival: Automatic vectorization of detections to MongoDB Atlas.
Semantic Search: Visual similarity and Natural Language search via Vertex AI Multi-modal embeddings.
Vertex AI Integration: Currently experiencing issues; the system gracefully degrades to Local-Only Mode for stable demonstrations.

🔌 Running the Demonstration

1. Start the FastAPI Backend

The API server must be running to provide the NMS-Free inference logic.

python main.py

2. Launch the Operations Dashboard

In a separate terminal, run the Streamlit UI:

streamlit run demo_app.py

3. Local Data Contract

Every batch job generates a manifest.json in batch_outputs/. This file demonstrates the "Subscriber Payload"—a clean JSON structure ready for hand-off to corporate ERP systems or future vector databases.

🏗️ Architecture (DDD)

The prototype follows Domain-Driven Design to ensure that while this is a PoC, the underlying logic is built on a scalable enterprise foundation:

src/api/: Model orchestration and API endpoints.
src/application/: Industrial use cases (Batch, Analysis).
src/domain/: Core PalmOilBunch business entities.
src/infrastructure/: Repository logic and AI Service wrappers.

README.md 4.3 KB Histórico Raw