Mauricio González Ortiz

About

Work

1. Parvada

   Software and deployment engineer, AI/ML

   Feb 2025 to May 2026, Clovis, NM

   • Eliminated manual chicken classification with an end-to-end AI pipeline in Python and C++ that processes 3000 chicks a minute: YOLOv8 detection, wing localization and a ViT classifier, wired to PCIe actuators for automatic separation.
   • Cut transportation costs by 50% with a local AI diagnostics agent (OpenAI when online, Llama when offline) that tells software failures from hardware ones, applies software fixes and pinpoints the parts that need maintenance.
   • Cut inference latency by 33% by implementing grayscale inference in Ultralytics before it was supported natively.
   • Raised model accuracy by 60% with a high-performance C++ machine vision library for an industrial high-speed camera. Frames go from capture to pinned memory to GPU inference with asynchronous, zero-copy transfers, so each chicken is seen in more frames.
   • Designed and built a custom C++ object tracker that re-identifies object IDs 10× faster than conventional tracking algorithms.
   • Grew clients by 30% by shipping a TypeScript and Node.js mobile app with hybrid CNN inference (on-device when offline, an Amazon SageMaker endpoint when connected) that classifies chickens by sex and counts them in a photo.
   • Cut client operating costs by 50% by owning the full deployment lifecycle (hardware assembly, model training and integration, machine validation and post-launch monitoring) as the sole technical contact for three live production systems.

2. Okus AI

   Data scientist and ML engineer

   Feb 2024 to Mar 2025, remote (USA)

   • Doubled production throughput with an end-to-end ELT pipeline that pulled raw video from several factory areas, ran CNN pose detection and segmentation, and turned it into structured production records: process, workers, machine ID, start and end times and other operational metrics.
   • Doubled the client base by building the data infrastructure that stores those records in Amazon DynamoDB, and integrating an Amazon Bedrock LLM agent that analyzes them and writes a full operations report.
   • Built a Python pipeline that ingested more than 50,000 ClinicalTrials records and used NLP entity extraction and market-sizing models to rank fields by their potential for AI adoption.

3. Harvard Medical School

   Undergraduate research fellow, Zhang Lab (software and computation)

   Feb 2022 to Feb 2023, Boston, MA

   • Cut manual microscopy annotation time by about 60% with Python image-processing pipelines that analyze cell distribution automatically, relating cell size to population frequency.
   • Ran finite element simulations of soft magnetic materials in FORTRAN and ABAQUS, modeling material behavior and physical interactions.

Skills

AI and LLMs

LangChain, LangGraph, RAG, multi-agent systems, generative AI, prompt engineering and versioning, tool calling, OpenAI SDK, Anthropic SDK, Ollama, Hugging Face

ML and vision

PyTorch, YOLOv8 and Ultralytics, computer vision, deep learning, NLP, model training and fine-tuning

Evaluation

precision@k, MRR, BLEU, perplexity, drift detection, retrieval evaluation, regression testing

Retrieval

OpenSearch (AWS), ChromaDB, Pinecone, vector indexing, structured extraction, chunking strategies

Backend

Python, FastAPI, TypeScript, Node.js, C++, Bash, SQL, REST and webhook design, CI/CD, pytest

Infrastructure

Linux administration, Docker, ONNX, distributed systems, AWS (EC2, S3, endpoints), Git

Languages

Spanish (native), English (bilingual)

Contact