Ernest H.

Key Skills

Software

Other

Top Subject Matter

Hospitality, F&B & Commercial Kitchen Operations

Facility Management & Maintenance (incl. HVAC)

Multilingual Customer Support (English / Mandarin)

Top Data Types

Image

Text

Document

Top Task Types

Classification

Question Answering

Evaluation/Rating

RLHF

Prompt + Response Writing (SFT)

Function Calling

Data Collection

Text Generation

Freelancer Overview

Self-taught AI builder and operations leader. Independently designed and shipped a production multilingual WhatsApp support bot serving licensees across 12 facilities and 180+ commercial kitchens in Singapore and Australia — owning the full pipeline: a 16-category issue classification taxonomy, few-shot example curation, multi-model routing (Claude Sonnet 4.5, Groq LLaMA 3.3), and ongoing output evaluation against real user conversations. Also built a vision-AI workflow for diagnosing air-conditioning control panels from real-world equipment photos. Bilingual (English / Mandarin Chinese) with deep domain expertise in F&B operations and commercial kitchen workflows.

Labeling Experience

Internal Knowledge Bot

TextQuestion Answering

Built an internal AI WhatsApp chatbot enabling operations staff to self-serve technical documents — floor plans, kitchens dimensions, equipment specs, and vendor SOPs — across 10 sites and ~250 commercial kitchens in Singapore. Scope and labeling tasks: Structured and tagged the entire technical document library by site, kitchen unit, document type, and content category for retrieval-augmented generation. Curated question/answer pairs from real operations-staff queries (floor-plan lookups, equipment dimensions, LPG and safety procedures) to validate retrieval accuracy. Defined query-intent categories — floor plan retrieval, equipment specs, vendor SOPs, safety/compliance lookups — to route questions to the right document set. Reviewed and labeled retrieval failures to refine document chunking, intent matching, and response phrasing. Project size: 10 sites, ~250 commercial kitchens, several hundred indexed technical documents. Quality measures: Service-account access control restricting bot use to registered staff only. Retrieval accuracy spot-checked against ground-truth source documents. Question/answer pairs reviewed for staff-language clarity in English and Mandarin.

2026 - Present

Vision AI Evaluation — Commercial HVAC Diagnostics

ImageClassification

Built a vision-AI workflow for diagnosing air-conditioning control panels in commercial kitchen facilities, where licensees frequently report unfamiliar error codes. Scope and labeling tasks: Curated a reference set of real-world commercial AC control panel photos captured across 12 facilities in Singapore and Australia. Labeled each image with the displayed error code, the affected component, the suggested technician action, and a customer friendly response template. Tagged failure modes where vision models misread codes due to glare, partial visibility, or unfamiliar panel-model variants — used to refine prompt instructions and routing rules. Project size: 12 facilities, 180+ kitchens. Dataset grows continuously as new licensee submissions arrive. Quality measures: Each model interpretation cross-checked against ground-truth error codes and actual maintenance outcomes before being added to the reference set. Multilingual response templates (English / Mandarin) reviewed for clarity before deployment.

2026 - Present

Multilingual AI Support Bot — Classification & Output Evaluation

TextClassification

Sole builder and data curator for an AI-powered WhatsApp support bot deployed across 12 facilities and 180+ commercial kitchens in Singapore Key labeling and evaluation work: Designed a 16-category classification taxonomy for facility and kitchen support requests, including urgency levels and emergency escalation criteria. Curated bilingual (English / Mandarin) few-shot examples used by an LLM router to route messages between models. Iteratively evaluated and rated outputs from Claude Sonnet 4.5 and Groq LLaMA 3.3 against real licensee conversations. Reviewed and labeled edge-case messages to refine prompt behavior and reduce misclassification.

2026 - Present

Education

Work History