Percy Cubas | ML Engineer

About

AI Engineer with experience spanning computer vision, deep learning, robotics, and full-stack development. I build end-to-end ML pipelines, from data acquisition and model training to cloud deployment, and enjoy bridging the gap between research and production systems.

Python C++ TypeScript PyTorch OpenCV YOLO Mask R-CNN ONNX W&B LangGraph LangSmith ROS-Noetic Gazebo ZED SDK Jetson Xavier AWS Docker Git Linux

Publications

4 Scopus-indexed (INTERCON x2, ANDESCON x2) + 1 IROS workshop

Deep Learning Approach for Accurate Pre-Harvest Blueberry Ripeness Classification

Workshop at IEEE/RSJ IROS, Oct 2023

Artificial Vision Strategy for Ripeness Assessment of Blueberries

P. Cubas, R. Huaman, S. Prado

IEEE INTERCON, Sept 2023

Deep Learning-based Segmentation and Classification System for Artichoke Seedling Grading

P. Cubas, R. Huaman, S. Prado

IEEE INTERCON, Sept 2023

Detection and Classification of Ventura-Blueberries in Five Levels of Ripeness

P. Cubas, E. Fiestas, S. Prado

IEEE ANDESCON, Nov 2022

Paper

IIoT System for Monitoring and Analysis of the Transplanting Process of the Artichoke Seedling

P. Cubas, J. Alva, S. Prado

IEEE ANDESCON, Nov 2022

Paper

Experience

Industry

Indra

  • Contributed to system design, technical coordination, and development of multiple AI projects
  • Escania: Full-stack document defect detection platform with YOLO pipeline, SageMaker training, and serverless inference
  • Ragia: Highly scalable Agentic RAG platform with hybrid retrieval, semantic chunking, and autonomous tool-calling agents
  • Analitia: Automated QA system for call centers using AWS Transcribe and parallel AI agents

YaVendio

  • Implemented AI-Agents for automated sales using LLMs (OpenAI, Claude, Gemini), LangSmith, and prompt engineering
  • Developed scripts and a CLI toolkit to streamline internal processes with MCP integration in Cursor
  • Improved LangGraph-based prompt generation API, optimizing interaction quality and AI-Agent adaptability

IDEGO STANDARD

  • Maintained and added features to Maersk Peru's order management platform, including import/export flows and document handling
  • Implemented backend microservices in Python/Django integrated with Azure services
  • Led a team of 3 developers expanding the financial back office for BCP account transfers

Research

LABINM - UPAO, Trujillo, Peru

  • Contributed to PROCIENCIA-funded project (PE501086701-2024) in precision agriculture and robotics
  • Built a data acquisition platform on ROS for a mobile robot equipped with a ZED 2i stereo camera, enabling non-technical users to capture video from field environments
  • Led a team of 3 interns building a modular pipeline for preprocessing and labeling
  • Managed code versioning and collaboration using GitHub across training and labeling workflows

LABINM - UPAO, Trujillo, Peru

  • Designed and implemented detection and segmentation strategies for crops (blueberries and artichokes) using YOLOv7/v8, Mask R-CNN, and MLPs
  • Configured simulation environments with ROS-Noetic, rviz, Gazebo, and ZED cameras, integrating models on a mobile robot with Jetson Xavier
  • Collaborated on the redesign of a PPG-based optosensor, implementing digital filters in C++ for Arduino

LABINM - UPAO, Trujillo, Peru

  • Restored and deployed an IIoT system for an artichoke project, configuring and managing the Linux environment
  • Researched state of the art in crop detection and classification, evaluating algorithms such as Mask R-CNN and YOLO
  • Managed blueberry image datasets, including labeling with Roboflow

Projects

LabelFlow: Video Annotation Tool

Open-source video annotation tool with motion-aware propagation using optical flow, YOLO re-detection, and region matching. Supports automatic object tracking across frames with persistent identity assignment and bulk operations.

Optical Flow YOLO Object Tracking

Repository

Signature Verification via Siamese CNNs

Siamese CNN for signature comparison based on learned representations, deployed as API on AWS (Lambda, EC2). Optimized inference using ONNX runtime and tracked experiments with Weights & Biases.

Siamese CNN ONNX AWS Lambda W&B

Repository

ML Models Review

Implementation of state-of-the-art Deep Learning models from scratch using PyTorch, focused on image classification and semantic segmentation.

PyTorch Classification Segmentation

Repository

Education

B.S. in Electronic Engineering

July 2023 Universidad Privada Antenor Orrego, Trujillo, Peru