Machine Learning · Computer Vision · Robotics

Naveen Prakash R

I'm a Machine Learning engineer building perception and intelligence for the physical world: self-supervised foundation models that guide autonomous robots, agentic LLM systems, and vision transformers optimised to run on the edge.

Based in Ireland Stamp 1G, eligible to work full-time in Ireland MSc Intelligent Systems, Trinity College Dublin
01

About

I'm a Machine Learning and Computer Vision engineer with four years of experience taking models from research to production, across robotics, medical AI, radar perception, and document automation. I like the hard parts: making a vision transformer run in real time on a Jetson, lifting 2D keypoints into 3D gait biomarkers, or wiring an LLM into a loop of real-world data tools.

Crucially, I ship models rather than leaving them in notebooks. I've built end-to-end production pipelines on AWS, training and serving on SageMaker, packaging inference in Docker with images versioned in ECR, automating build, test and deployment through CI/CD (Jenkins), tracking experiments with MLflow, and provisioning infrastructure with Terraform. I'm just as comfortable shipping to the edge, converting and optimising models to ONNX for real-time inference on NVIDIA Jetson hardware, and serving models behind REST and FastAPI APIs.

I'm currently a Research Assistant at the University of Galway, building autonomous navigation for a quadruped robot using self-supervised foundation models for search-and-rescue. I hold an MSc in Computer Science (Intelligent Systems) from Trinity College Dublin and I'm an AWS Certified Data Engineer.

4+ Years building ML systems
2 Patents filed
1 IEEE publication
AWS Production ML on cloud & edge
02

Experience

  1. Research Assistant · School of Computer Science, University of Galway

    Jun 2026 – Present
    • Designed and built TerraBase, a real-time autonomous navigation platform (ROS 2 Humble, Linux) for a Unitree Go2 quadruped, integrating a self-supervised DINOv2-based visual-traversability foundation model to flag unfamiliar terrain in unseen environments, relevant to search-and-rescue.
    • Deployed and optimised a DINOv2 Vision Transformer for real-time on-device inference by converting it to ONNX for an NVIDIA Jetson Orin NX (JetPack 6.2, CUDA 12.6, PyTorch), cutting per-frame latency to fit the embedded edge compute budget.
    • Integrated the Jetson compute unit with the physical Go2 robot and a ZED 2i stereo camera, improving real-time foundation-model performance on edge hardware.
    ROS 2 HumbleDINOv2Vision TransformersJetson Orin NXONNXCUDAOptical Flow
  2. Machine Learning Engineer · SiPra (Remote)

    Oct 2024 – Nov 2025
    • Designed a deep-learning pipeline to predict early Alzheimer's risk from gait patterns using 2D and 3D skeletal keypoints across 100+ subjects and multiple walking variations.
    • Built a multi-stage architecture combining YOLO + OpenPose pose extraction, Vision Transformer feature encoding, relative spatial encoding, and LSTM temporal modelling for 3D gait-trajectory analysis.
    • Modelled spatiotemporal gait dynamics by lifting 2D keypoints to 3D representations and learning sequence-level biomarkers correlated with cognitive decline.
    Computer VisionYOLOOpenPoseViTLSTMPyTorch
  3. Machine Learning Engineer · MulticoreWare

    Dec 2021 – Apr 2023
    • Designed modular ML inference services with REST APIs, enabling integration of radar-based activity-recognition models into downstream applications.
    • Implemented CI/CD workflows with Jenkins to automate compilation, testing, and deployment of PyTorch operator optimisations for accelerator-based inference.
    • Built a radar data pipeline with Airflow, applying FFT feature extraction and DBSCAN clustering to feed BlazePose + LSTM activity/identity models reaching up to 95% accuracy.
    • Built a production ML deployment pipeline on AWS, training and serving via SageMaker endpoints, Docker images in ECR, experiment tracking with MLflow, and infrastructure provisioned with Terraform.
    AWS SageMakerDockerTerraformAirflowMLflowJenkinsC++DBSCAN
  4. AI Developer · Kyrios Ark World Logistics (Freelance)

    Dec 2020 – Dec 2021
    • Developed an end-to-end OCR + NLP automation pipeline (Tesseract + spaCy) to extract invoice metadata (HS codes, quantities, CIF values) from scanned documents, reducing human error by 70%.
    • Deployed a WhatsApp shipment-tracking bot with Twilio + Flask, letting clients query real-time container status in natural language, backed by automated Google Sheets updates.
    • Designed a truck ETA prediction system using the Google Maps API and traffic-aware logic, improving dispatch accuracy and reducing missed port cut-offs.
    Tesseract OCRspaCyNLPFlaskTwilio APITensorFlow

Education

Sep 2023 – Sep 2024 MSc Computer Science (Intelligent Systems) Trinity College Dublin

Thesis: Burst Image Denoising Using Optical Flow

  • Developed an optical-flow-based advanced burst image denoising approach.
  • Enhanced image alignment and denoising by improving the fusion of multiple frames under challenging conditions such as low light and motion blur.
  • Targeted mobile photography and surveillance applications, delivering higher image clarity.
Computer VisionPythonOptical FlowImage Processing
Jun 2017 – Jun 2021 BE Electronics & Communication Engineering SRM University

Thesis: Automated Inventory and Order Assistance System

  • Designed a computer-vision and IoT system for automated inventory tracking and order assistance, published as an IEEE paper.
Computer VisionPythonYOLOv5Image ProcessingSensor NetworksIoT
03

Featured Projects

Agentic LLM System

Ireland Manufacturing Site-Selection Agent (ITag)

A multi-tool agentic LLM system built on the Anthropic Claude API with a custom orchestration loop. It parses natural-language manufacturing requirements into structured SiteRequirements objects, runs six concurrent data-fetching tools over real IDA Ireland, CSO, EPA and SEAI datasets, then ranks counties with a weighted composite scoring engine and exports top-5 site recommendations with per-factor breakdowns and risk annotations.

Anthropic Claude APIAgentic OrchestrationTool UseParallel ExecutionConstraint ParsingPython
View code ↗
LLM Fine-Tuning

LLaMA 3.2 Fine-Tuning for Text-to-SQL

Fine-tuned LLaMA 3.2 (3B Instruct) with LoRA / PEFT to convert natural-language queries into accurate SQL while minimising trainable parameters and GPU memory. Built a GPU-accelerated inference pipeline on Kaggle with HuggingFace Transformers and torchao (quantisation + optimised loading), with model versioning via the HF Hub and reproducible LoRA weights packaged as distributable datasets.

PyTorchHuggingFacePEFT / LoRAtorchaoQuantisationMLflow
Robotics & Edge AI

TerraBase: Autonomous Quadruped Navigation

Real-time autonomous navigation for a Unitree Go2 robot (ROS 2 Humble) using a self-supervised DINOv2 visual-traversability model to detect unfamiliar terrain for search-and-rescue. The Vision Transformer is exported to ONNX and optimised to run within the latency budget of an NVIDIA Jetson Orin NX, paired with a ZED 2i stereo camera.

ROS 2DINOv2Edge AIONNXJetson
Embedded Vision

Smart Accident-Prevention Helmet (Patented)

A patented rider-safety helmet integrating bone-conduction audio and a heads-up display, with an embedded NVIDIA Jetson Nano running a real-time traffic-sign and signal recognition model for on-device inference and rider alerts.

Patent 202141060755Jetson NanoEdge InferenceObject Recognition
04

Skills & Toolbox

Languages

PythonC++CSQLPySparkReactJS

Deep Learning & CV

PyTorchTensorFlowOpenCVVision TransformersLSTMYOLOImage ProcessingQuantization

LLM & GenAI

LangChainLoRA / PEFTAgentic SystemsAnthropic Claude APIHuggingFaceStreamlit

MLOps & Production

AWSSageMakerDockerKubernetesTerraformCI/CD (Jenkins)FastAPIREST APIsMLflowAirflowGit

Robotics & Edge

ROS 2Jetson Orin / NanoRaspberry PiArduinoESP32ONNX

Data & ML

Scikit-LearnNumPyPandasMatplotlibSeabornMongoDBDjangoFlask
05

Achievements & Credentials

AWS Certified Data Engineer, Associate (DEA-C01)

Designed secure, scalable batch & streaming pipelines with S3, Glue, EMR, Kinesis and Redshift, applying data-modelling, partitioning and performance best practices.

IEEE Publication

Automated Inventory Management System, published in IEEE Xplore. Read the paper ↗

Two Patents Filed

Smart accident-prevention helmet (App. 202141060755) & an automated smart-kitchen robot for making stuffed parathas (App. 202141060759).

Smart India Hackathon 2020: Finalist

Selected as a national finalist in the Hardware Edition of India's flagship innovation hackathon.

06

Let's build something

I'm open to ML / Computer-Vision and AI-engineering roles in Ireland and remote. The fastest way to reach me is email.