Samuel Hai

Key Skills

Software

Clickworker

Data Annotation Tech

Google Cloud Vertex AI

Labelbox

Mercor

OneForma

Other

Top Subject Matter

No subject matter listed

Top Data Types

Audio

Image

Text

Top Task Types

Bounding Box

Classification

Object Detection

Question Answering

Text Summarization

Freelancer Overview

I started training AI 4 years ago using what is now Google's Vertex AI. The first major project was an person detection project whereby I was training a computer vision model on detecting the presence of people within forklift zones. To seed the data, I used Lego people and small RC cars, but as time went on, I was able to use real people, real caution tape, and real heavy equipment. Along similar lines, I also did a person detection project for restaurants, where I trained computer vision to detect if a table was open/occupied and messy/clean. My next training project was for a retail client. The requirement was to leverage existing store security cameras to detect when products are misplaced, damaged, or not on the shelf (but still in stock). I couldn't use a real location for this, but my company bought some grocery store shelves as well as a variety of products so that I could arrange them similarly to a grocery aisle. I then took a few thousand photos of common loss-type scenarios, like open box, product out of place, empty section on shelf, and even attempts to change price stickers. This was eventually taken over by another team that specialized in retail technology consulting. On the side I also annotate on Multimango, ClickWorker, and Mercor

Labeling Experience

Response Evaluation

OtherTextClassificationText Summarization

I would be given a prompt, and I would need to identify the prompt's language as well as the type of prompt. Then I would need to evaluate 2 responses from different models on instruction following, format, accuracy, helpfulness, and overall quality.

I would be given a prompt, and I would need to identify the prompt's language as well as the type of prompt. Then I would need to evaluate 2 responses from different models on instruction following, format, accuracy, helpfulness, and overall quality.

2025

Chinese Transcription

OtherAudioText GenerationTranslation Localization

I would hear Chinese audio clips, and evaluate the accuracy of a model's transcription, making edits where needed

I would hear Chinese audio clips, and evaluate the accuracy of a model's transcription, making edits where needed

2025

GUI/Console

OtherImageBounding Box

I was tasked with collecting screenshots of various computer software, then I needed to place bounding boxes on specific buttons/functions, or identify if the button/function wasn't there.

I was tasked with collecting screenshots of various computer software, then I needed to place bounding boxes on specific buttons/functions, or identify if the button/function wasn't there.

2025

Grocery Shelf

Google Cloud Vertex AIImageObject Detection

The requirement was to leverage existing store security cameras to detect when products are misplaced, damaged, or not on the shelf (but still in stock). I couldn't use a real location for this, but my company bought some grocery store shelves as well as a variety of products so that I could arrange them similarly to a grocery aisle. I then took a few thousand photos of common loss-type scenarios, like open box, product out of place, empty section on shelf, and even attempts to change price stickers

The requirement was to leverage existing store security cameras to detect when products are misplaced, damaged, or not on the shelf (but still in stock). I couldn't use a real location for this, but my company bought some grocery store shelves as well as a variety of products so that I could arrange them similarly to a grocery aisle. I then took a few thousand photos of common loss-type scenarios, like open box, product out of place, empty section on shelf, and even attempts to change price stickers

2022 - 2022

Forklift Zone

Google Cloud Vertex AIVideoBounding BoxObject Detection

This project involved training a model to use a warehouse's overhead cameras to detect the presence of people in forklift zones. Zones were boxed with bright caution tape. The model was to recognize if a person or other object was in any of the boxes, and to trigger an alert to whoever is responsible for onsite safety.

This project involved training a model to use a warehouse's overhead cameras to detect the presence of people in forklift zones. Zones were boxed with bright caution tape. The model was to recognize if a person or other object was in any of the boxes, and to trigger an alert to whoever is responsible for onsite safety.

2021 - 2021

Education

Work History