Skip to content
View levipereira's full-sized avatar
🎯
Focusing
🎯
Focusing
62 followers · 18 following

Achievements

Achievement: Pull Sharkx2Achievement: Starstruckx2

Achievements

Achievement: Pull Sharkx2Achievement: Starstruckx2

Highlights

Block or report levipereira

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
levipereira/README.md

Levi Pereira

MLOps Engineer & Computer Vision specialist focused on NVIDIA AI infrastructure — from model training to production deployment at scale.

5 years building end-to-end ML pipelines for real-time video analytics, backed by 14 years of enterprise infrastructure and database engineering. I care about systems that actually work in production: optimized, reliable, and maintainable.

What I work on

  • Object Detection & Tracking — YOLO family (v3 through v11, YOLO-NAS), custom training pipelines, SAHI for small object detection in high-resolution imagery
  • Model Optimization — TensorRT INT8/QAT quantization, ONNX export, latency/accuracy tradeoffs for production constraints
  • Video Analytics Infrastructure — NVIDIA DeepStream, GStreamer plugin development (C++/Python), multi-camera pipelines
  • MLOps — Model lifecycle management, drift detection, automated monitoring and alerting, CI/CD for ML systems
  • Edge & GPU Deployment — NVIDIA Jetson, Triton Inference Server, containerized inference stacks serving 1000+ concurrent requests

Contributions to NVIDIA DeepStream

Two features I proposed were accepted into the official DeepStream SDK:

  • objStatus in Python bindings — added objStatus field to pyds.NvDsAnalyticsObjInfo, merged into DeepStream's internal branch for the next release · thread
  • Runtime analytics config reload (DeepStream 8.0) — implemented runtime configuration reload for nvdsanalytics, REST API extensions for nvmultiurisrcbin, and nvds_rest_server integration · thread

Stack

Detection     YOLO v3/v4/v5/v6/v7/v8/v9/v10/v11 · YOLO-NAS · RT-DETR · D-FINE
Segmentation  SAM · SAM2 · SegFormer · DeepLab v3+
Transformers  ViT-based detectors · Grounding DINO · CLIP
Other CV      Pose Estimation · OCR / Document AI
Optimization  TensorRT · ONNX · INT8 QAT · CUDA · cuDNN · SAHI (sliced inference)
Serving       NVIDIA Triton Inference Server · DeepStream · GStreamer
MLOps         TAO Toolkit · model monitoring · drift detection · Kubernetes
Languages     Python · C++ · Bash · SQL
Infra         Docker · PostgreSQL · Kafka · FastAPI · Linux (AIX · RHEL · Ubuntu)
Databases     Oracle (RMAN · ASM · RAC) · PostgreSQL
Data          Database design · Schema modeling · Query optimization · PL/SQL

Background

Before AI/ML, I spent 14 years (2006–2020) as a Database Administrator and Infrastructure Specialist across AIX, Linux, and Windows environments — Oracle databases, high-availability clusters, storage, and enterprise ops. That background shapes how I approach production ML systems: reliability and operational discipline aren't afterthoughts.

Contact

Pinned Loading

  1. yolov9-qat yolov9-qat Public

    Implementation of YOLOv9 QAT optimized for deployment on TensorRT platforms.

    Python 140 18

  2. deepstream-yolo-e2e deepstream-yolo-e2e Public

    Implementation of End-to-End YOLO Models for DeepStream

    Python 76 9

  3. deepstream-sahi deepstream-sahi Public

    Native GStreamer plugins that integrate SAHI (Slicing Aided Hyper Inference) into NVIDIA DeepStream for real-time small object detection in high-resolution video streams.

    C++ 34 5

  4. deepstream-community-patches deepstream-community-patches Public

    1

  5. Model-Optimizer-ONNX Model-Optimizer-ONNX Public

    ONNX post-training quantization (PTQ) and TensorRT deployment helpers for YOLO-style detectors

    Python 5