Custom AI Solutions

From research questions, data collection and experimental conditions to analytical frameworks, we turn research ideas into methods that can be carried out and validated.

From data preparation, model training and adjustment to search, ranking, recommendation and performance evaluation, we choose methods to suit the problem at hand.

We build AI systems that can look up information, use tools, handle multi-step tasks and manage knowledge, including workflows, result checks and error handling.

We connect speech recognition, content understanding, generation and speech synthesis, and can design dedicated workflows for specific language, cultural or communication needs.

Depending on the project, we combine speech, text, images, wearable device data, physiological signals or other behavioural data to address longitudinal analysis, limited-data and cross-context analysis challenges.

From device control and hardware debugging to image recognition and edge deployment, we extend AI beyond software into real-world settings.

Team members have contributed to search, recommendation and ranking systems for large-scale everyday services and e-commerce platforms.

From a user expressing a need to a system finding relevant results in a large pool of information, prioritising them and improving continuously based on actual behaviour, the work involves candidate retrieval, ranking, model training, user behaviour analysis and system optimisation.

The team has also contributed to large language model post-training and real-world search applications, gaining experience in taking model capabilities into products operating at scale.

A team member contributed to a financial services accessibility project that converted ordinary financial text into structured representations for a Hong Kong Sign Language generation system.

The process involved more than asking a large model to “translate”. It incorporated terminology retrieval, grammatical transformation, rule checks, fallbacks and structured output validation, with repeated refinements alongside sign language experts.

This experience led the team to address an important question early on:

When general-purpose AI may not suit a particular group of users, how should the entire interaction workflow be redesigned?

A team member independently developed a complete voice AI workflow, covering speech recognition, LLM generation, speech synthesis, subtitle alignment and video output.

The system also incorporated evidence checking, low-confidence fallbacks and quality checks, making it possible to identify which step had gone wrong when a problem arose, rather than discovering only at the end that the result was unreliable.

This experience can also extend to voice AI and multi-step workflows with higher reliability requirements.

The team has research experience in intelligent hardware, robot control and computer vision.

Previous projects include remotely controlled robots and image recognition for passenger detection and occupancy estimation inside double-decker buses, addressing occlusion, lighting, motion blur and limited computing resources.

These experiences lead the team to consider not only whether a model is accurate, but whether a system can actually be used in the real world.

During the founder’s MPhil studies at the University of Cambridge, the team independently developed a conversational educational AI solution for schoolchildren in Hong Kong.

The project used English conversation practice as its setting, while exploring methods for observing learning patterns, engagement, metacognition and psychological states.

This experience also helped shape an important direction for the team today:

Conversation can do more than answer questions. It can offer a way to understand a person, collect longitudinal data and provide personalised interventions.