Things I've worked on

At City From Naught Studio, I led the initial product development of Marleen, a generative AI tool designed to create video game textures. In game development, achieving customized, high-quality textures traditionally required skilled artists proficient in specialized software. As a result, only game studios with substantial budgets could afford to use highly realistic game textures, while smaller studios often had to resort to off-the-shelf textures or settle for creating 2D or low-poly style games which doesn't require highly realistic textures. The product vision for Marleen is to provide indie studios with the ability to create triple A quality game textures with indie game budgets.

AI generated contents are fast, but has unpredictable outcome (the chance that the generated content is usable without any major tweaks is very low), while human artists are the opposite, where the process is slow but the results are predictable. Due to this difference, users often has to generate hundreds of results to find a usable result, or more often, give up if the first few results are not usable at all. So the greatest challenge is how to ensure users continue to use the product and guide them to produce a good result.

To differentiated Marleen from other similar products on the market, we decided to focused on three core principals:

• Ensure the initial generated results are of high enough quality that user see potential of improvement with further generation

• Allow user to provide feedback to the process with further description and generation guidance

• Each iteration should be better in quality than the next, until it ultimately arrive at a usable quality for the end user.

This allowed us to better retain users and ensure that Marleen deliver result vs just a cool product to try

Orbiseed's document extraction technology found a valuable application in the realm of plumbing product specifications. By harnessing this technology, we were able to streamline the process of determining which products to purchase for the construction projects based on engineering requirements. The manual process is time consuming and required a lot of attention to details, and the documents are often contradictory to each other, making it a difficult job that is very tedious and prone to error.

In collaboration with a plumbing specification company, we trained a machine learning model to extract product criteria from engineering drawings and specification documents. Leveraging this information, the model then queried a comprehensive database of plumbing products to generate a curated list of recommendations that aligned with the project's needs. Since accuracy is very important, we mixed multiple approaches to ensure that no data is missed. Then working closely with product specifiers to further ensure its accuracy and reduce false positives. 

The outcome was an AI-enhanced workflow that allows product specifiers to significantly speed up their specification process, improving their efficiency and accuracy.

Orbiseed's document extraction engine has proven to be beneficial not only in the construction industry but also in the insurance sector. Specifically, we identified its utility in the domain of risk assessment for insurance companies. Traditionally, risk engineers are responsible for evaluating factors such as building history and geological data, which underwriters rely on to determine appropriate insurance costs. This process often involves extensive manual reading of numerous documents, resulting in lengthy timelines that span several months.

To address this challenge, Orbiseed employed a Reinforcement Learning with Human Feedback (RLHF) approach. Through this methodology, we developed a machine learning model capable of extracting risk-related data with an impressive accuracy rate of over 90%. By automating the extraction process, our solution significantly reduces the time and effort required for risk assessment, enabling insurance companies to streamline their operations and make more informed underwriting decisions.

Another product we built at Orbiseed was a digital twin tool. It leverages data extracted from CAD files and engineering documents to offer a comprehensive 3D visualization of a building's systems. This includes a range of components such as fire protection systems, IoT device data, HVAC, electrical, plumbing, and more. Our primary audience consisted of building managers and real estate developers seeking enhanced management capabilities for their building's data systems.

Our second creation at Babylon VR is a remarkable virtual tour builder. This user-friendly tool empowers individuals to effortlessly generate and share immersive virtual tours using 360-degree photos. Our inspiration for this product stemmed from valuable customer feedback, which highlighted the desire for simplified sharing options and the flexibility to experience virtual tours without the need for a VR headset. Moreover, we expanded the scope beyond pre-built projects, allowing users to effectively market existing spaces through captivating virtual tours. Notably, building these experiences using 360 images significantly reduces costs compared to creating VR content using intricate 3D models.

At Babylon VR, we created an immersive tool that empowers home buyers to customize their living spaces by seamlessly integrating a VR headset and hand tracking device. Our target audience primarily consists of individuals interested in purchasing pre-built homes and condos. By leveraging our tool, they gain the freedom to explore and experiment with an array of furnishing options before making their final decisions, ensuring a personalized and satisfactory outcome.