Research Engineer — AI Agent Systems for On-Site Construction

Castor Innovations Limited; Hong Kong; On-site; HKD 900,000 - 1,300,000; Full time

Job summary

Castor Innovations Limited in Hong Kong is seeking a Research Engineer to push the core frontier of our agent platform. This hands-on role focuses on turning scientific uncertainty into production-grade reliability, with work centered in Hong Kong.

You will develop multi-agent orchestration, vision-language model reliability for on-site safety, Cantonese ASR, and a retrieval-augmented knowledge base, while evaluating closed-loop correctness.

Qualifications

• Education: a bachelor’s, master’s or PhD in Computer Science, AI, Electronic Engineering or related STEM field.
• Right to work: candidates have full working rights for Hong Kong for the full duration of the engagement.
• Core skills: strong Python and backend/systems engineering; hands-on with LLMs, RAG systems and/or multi-agent orchestration; depth in state-machine/distributed-systems, CV, or ASR.

Responsibilities

• Research determinism, auditability and production-grade reliability for multi-agent systems in construction/public works.
• Investigate reliability of vision-language models for on-site safety detection and PPE compliance.
• Develop Cantonese and multilingual ASR for noisy construction sites.
• Design the agent runtime including intent routing, state machines and data provenance.

Skills

Python

LLMs

RAG systems

Multi-agent orchestration

State-machine design

Vision systems

ASR

Education

Bachelor's/Master's/PhD in CS/AI/EE

Tools

PyTorch

TensorFlow

LangChain

ONNX

Job description

Castor Innovations Limited in Hong Kong is seeking a Research Engineer to push the core frontier of our agent platform. This hands-on role focuses on turning scientific uncertainty into production-grade reliability, with work centered in Hong Kong.

You will develop multi-agent orchestration, vision-language model reliability for on-site safety, Cantonese ASR, and a retrieval-augmented knowledge base, while evaluating closed-loop correctness.