ウェブファーマー List of Products and Services

In the medical system development project that our company is participating in, we are responsible for the research, development, and implementation of the AI components. To meet the precision and reliability required in the medical field, we are broadly engaged in the following tasks: ■ Research on relevant AI technologies ■ Design, development, and implementation of models ■ Training models using data ■ Verification of results and feedback Through these processes, we provide AI that supports the foundation of systems contributing to healthcare. We hope that many lives will be saved through our AI development. *For more details, please refer to the related links or feel free to contact us.*

We have implemented improvements to the image generation AI model using the Diffusion model operated by Company E. The Diffusion model runs the model through multiple steps during inference, which generally results in longer inference times. This is one factor that can diminish the user experience. Our company has a deep understanding of model architecture developed over many years, as well as an understanding of the device, which has allowed us to achieve faster and lighter inference without compromising accuracy. (Image is from DDIM) *For more details, please refer to the related links or feel free to contact us.*

In June 2024, we created an LLM for a certain company. In large corporations, the internal seeds have grown too large, making it difficult to generate new business ideas from them. This time, we developed a system where the LLM proposes new business ideas using the company's information A and another piece of information B. We used the open-source Llama model as the LLM. This allows the company to complete the work on-premises without disclosing information externally. To efficiently propose ideas, the system incorporates features like RAG and prompt modification. *For more details, please refer to the related links or feel free to contact us.*

In March 2024, we performed optimization to adapt an AI model for a specific edge device for semiconductor-related company M. We implemented lightweight modifications to adapt it to a special device. This time, the device for inference required significant lightweight adjustments in terms of model size, so we started by reviewing the neural network architecture. After that, we performed quantization and pruning to make it adaptable to the specific device. As a result, we achieved a model that meets the speed and data capacity requirements for the specific device. *For more details, please refer to the related links or feel free to contact us.*

In September 2022, we developed a system to promote DX for construction companies, as well as an AI system to read specific characters in business operations. This included everything from creating annotation tools to inference systems. We handled the creation of annotation tools, annotation, research, model implementation, training and evaluation, and the creation of the inference system. For the character recognition AI, we utilized an OCR service for part of the process, but we conducted preprocessing and correction using our proprietary AI before and after that. *For more details, please refer to the related links or feel free to contact us.*

In 2022, we conducted analysis on information obtained from domestic companies' biological devices. From research on time-series data to the construction of an inference system. This time, since the subject was time-series data, we selected journals and conferences that corresponded to it and chose an appropriate Neural Network model architecture from those papers. Furthermore, we implemented the model, conducted training and evaluation, built the inference system, and visualized the results. Visualizing the results. Time-series data is difficult to visualize, but without appropriate visualization, evaluating the results becomes challenging. This time, we employed several methods to creatively visualize the data. *For more details, please refer to the related links or feel free to contact us.*

In July 2022, we conducted research and development of an inference system using chemical sensors for domestic companies. From research to iPhone app development. This time, we started with literature research, followed by model implementation, training and evaluation, and the construction of the inference system, culminating in the app development for iPhone. We built a logic based on chemical research. This time, we dealt with substances in the AI industry that are considered a specialized field, specifically chemical sensors. Since there were no precedents for the inference model based on the output values, we thoroughly researched chemical literature and implemented a model that is chemically valid and highly explainable. We switched from Dart to Swift. Initially, the iPhone app was developed using the Dart programming language and Flutter. However, it became clear that some essential features could not be realized with Flutter, so we quickly switched to Swift. This allowed us to implement all the required features in the app. *For more details, please refer to the related links or feel free to contact us.*

From 2021 to 2023, we conducted research and development on an AI system to be integrated into home appliances for Company I. We developed everything from annotation tool creation to inference system construction in a consistent manner. This included creating the annotation tool, performing annotations, conducting research, implementing models, training and evaluation, model optimization, building the inference system, and converting it for edge devices. We conducted research and implementation tailored to a specialized neural network architecture. The current neural network processes video and time-series data as input, so we researched models that accommodate this and implemented them based on past examples. We visualized the inference results. Since this development was at the POC stage, we sent the internal inference results to another computer and displayed the results in an easily understandable manner. This made debugging and improvements easier. *For more details, please refer to the related links or feel free to contact us.*

From 2019 to 2023, we conducted research and development on an AI system to be integrated into beauty devices for Company H in Japan. This time, since there were no usable existing data, we started from data generation. We created an annotation tool and performed the annotation. Furthermore, we conducted research, implemented models, trained and evaluated them, reduced model size, and built an inference system. Due to the operational requirements of the hospitals using the devices, the allowed inference time per image was a few milliseconds. To achieve this, we carefully considered the architecture of the Neural Network and assembled an architecture that is ultra-fast yet reasonably accurate. The Neural Network architecture itself was made lightweight, but further reduction in size was necessary to meet the inference time on edge devices. Therefore, during the deployment phase, we performed pruning, distillation, and quantization. This allowed us to meet the required inference time. *For more details, please refer to the related links or feel free to contact us.*

From 2018 to 2020, I conducted detection, tracking, position estimation, and behavior estimation of animals using AI. I developed everything from the creation of annotation tools to the construction and visualization of inference systems. At that time, detection, tracking, and behavior estimation of animals were fields where research was not advanced, so I started from creating annotation tools and performing annotations. Furthermore, I conducted research, implemented models, trained and evaluated them, and built and visualized inference systems. *For more details, please refer to the related links or feel free to contact us.*

In April 2017, we researched and developed a model to predict real estate prices from structured data for real estate companies. From research to inference model creation. This time, we started with research that accommodates structured data and time series data, and proceeded to model implementation, training and evaluation, and the creation of an inference system. *For more details, please refer to the related links or feel free to contact us.*

In 2018, we built an anomaly detection system for the production line of a food factory. In food production environments, the presence of impurities and unsanitary substances can become problematic. Therefore, a high recall inference system is required. From searching for appropriate models to implementation, training, and building the inference system, we researched suitable models and carried out their implementation, training, evaluation, and the construction of the inference system. To avoid existing patents, we developed a new model. Many companies are working on anomaly detection in this type of production line, and this time it was necessary to construct a new model that does not infringe on already filed patents. Thus, we conducted research at CVPR and ICCV and implemented a newly proposed type of anomaly detection model. *For more details, please refer to the related links or feel free to contact us.*

In 2019, we developed a system for detecting and tracking specific substances in medical CT images for Company M. This involved creating an annotation tool, conducting research, implementing models, training and evaluation, and building an inference system. We created a proprietary tool. To meet the client's requirements, we needed special data, but there was no tool available for labeling it. Therefore, we independently developed an annotation tool and generated the data. Detection and tracking of special objects in unclear CT videos. In this case, the CT images in each frame of the video were coarse, making the detection and tracking of specific substances extremely challenging. However, we overcame this through improvements in the neural network architecture, innovations in image preprocessing, and enhancements to the tracking algorithm. *For more details, please refer to the related links or feel free to contact us.*

In 2025, a system using AI agents was developed to streamline the operations of Company E. The AI agent understands the situation and plans the tools to be used. Based on this plan, inquiries to the LLM and RAG are executed. Both are repeatedly utilized based on the results. This process continues until the final output meets the required specifications. Programming language used: Python 3 Libraries used: Langgraph, Langchain Cloud: Azure *For more details, please refer to the related links or feel free to contact us.*