ANGEWORK List of News Articles

ANGEWORK Co., Ltd. has announced a system called "3D Zankurundesu" that converts spaces into 3D from photos and videos. Using data captured with smartphones and digital cameras, it generates a 3D space that can be viewed from different perspectives after shooting. It does not require an internet connection at the time of shooting or a dedicated 3D scanner, allowing the use of existing shooting equipment. Targeting construction and renovation, facility and real estate management, factories and warehouses, and photography businesses, it supports PoC and system development for the introduction of 3D recording, integration of 3D display into existing systems, and shooting support functions tailored to business needs. As a proof of concept, a stone wall at the A6 exit of Jimbocho Station, before its closure, was photographed for about 20 minutes with an iPhone 16 Pro and published as 3D data viewable in a web browser. For corporations considering implementation, a one-time free preliminary verification using provided photos and videos will also be conducted. For detailed features, applications, and shooting conditions, please refer to the related page.

ANGEWORK Co., Ltd. participated in the development of the Windows version of the powder diffraction data analysis software "Z-Rietveld," developed by Project Z-Code. We are responsible for the Windows version GUI using .NET 8 WPF, the C++ core integration API bridge, the design and implementation of the GUI structure including histogram display, the installer, and the build environment setup. Z-Rietveld is advancing the commonality of the core between the Macintosh and Windows versions, refreshing it into a framework that shares input/output processing and analysis routines. We would like to introduce our development achievements to those interested in Windows compatibility, porting, GUI development, and API integration for specialized analysis software used in universities, research institutions, and manufacturing industries.

ANGEWORK Co., Ltd. used the Raspberry Pi 3 Model B (RAM 1GB) from about 10 years ago to implement and optimize generative AI and verify the operation of small LLMs, publishing two technical articles. In the porting of small AI for ESP32, the generation speed was accelerated approximately 6.3 times from 31.22 tok/s to 195.81 tok/s through software optimization. Furthermore, multiple small LLMs were compared, and a model capable of Japanese conversation was confirmed. The approach to design, model selection, and optimization to establish AI in limited CPU and memory environments is introduced, rather than relying solely on expensive GPUs and the latest equipment. This content is aimed at corporate representatives and engineers considering AI utilization in existing devices, IoT, manufacturing equipment, and closed environments. For more details, please refer to the note article.

ANGEWORK Inc. has released a contract development package called "LocaPhone," which integrates local LLM (on-device AI) into existing iOS applications, on Lancers. By completing AI inference within the iPhone, it ensures that personal and business information is not transmitted externally and can be used in environments without communication. This setup is suitable for specific implementations such as reception and facility guidance, standard FAQs, organizing input text, short text summarization, and in-game NPCs. Our company has confirmed the complete offline operation, including Japanese responses, using Gemma 4 E2B on the iPhone 16 Pro. We will verify the target devices, existing applications, and use cases, and will respond gradually from PoC to actual device testing.

ANGEWORK Co., Ltd. has implemented the latest small LLM "LFM2.5-2.6B" released by Liquid AI on the iPhone and published the results of the actual device verification on note. In this verification, summaries, reasoning, Japanese responses, JSON format output, and hallucination suppression were evaluated under the same conditions, and compared with the existing small LLM "Gemma" to confirm their respective features and practicality. The article also discusses the point that performance cannot be judged solely by model size, as well as the considerations for selecting small LLMs in on-device AI development. This information is published as technical information for companies and developers considering local LLMs and AI systems for smartphones.

ANGEWORK Co., Ltd. has published an article titled "Nostalgic Game Book! How to Choose the Right AI for You - August 2026 Edition," in response to voices expressing interest in AI but finding it "overwhelming and difficult to understand due to the many types available." The article explains the applications and selection of AI in a game book format. Our company representative, who is involved in IT development and AI utilization, has structured the content to allow readers to learn by following options based on their goals, such as "I want answers to questions," "I want to create images or text," "I want to research documents," or "I want to operate existing applications," while minimizing the use of technical jargon. The article broadly introduces the main applications and mechanisms of current AI, including generative AI, image, video, audio, code, 3D generation, internal document search, local AI, AI agents, MCP, skills, and harnesses. It is designed to serve as an entry point for AI selection, applicable for personal use as well as business implementation.

ANGEWORK Inc. has successfully integrated Google's lightweight LLM "Gemma 4 E2B" into the iPhone 16 Pro, enabling on-device AI to generate Japanese responses without internet communication. In this verification, we implemented a conversation screen using SwiftUI, bundled the model within the app, addressed crashes during device startup, adjusted Japanese input, and enabled local storage of conversation history. We confirmed functionalities such as short Japanese conversations, rephrasing of business texts, and generating item lists. Since it does not send information to the cloud, it is expected to be utilized in fields requiring confidentiality, such as healthcare, manufacturing, and research, as well as in facilities without communication environments, and for reception, guidance, and input assistance. Moving forward, we will focus on improving speed, reducing memory usage, and enhancing integration with closed-domain RAG and existing applications.

ANGEWORK Inc. (Headquarters: Chiyoda, Tokyo; CEO: Yuhiro Kiryu) released the waveform data extraction and analysis engine "Cherubim Wave Core" on July 9, 2026. Cherubim Wave Core is a software engine that extracts waveform information displayed or output by measuring instruments and dedicated applications, allowing it to be utilized as data in formats such as CSV, TXT, and xlsx. It supports the digitization and analysis of waveform data presented on paper, screens, or in image form without replacing existing measuring instruments, facilitating data collection for research, statistical analysis, and AI applications. It is compatible with various waveform display devices such as oscilloscopes, material testing machines, EEGs, and ECGs, and can also accommodate customized development based on the usage environment. Initial consultations are free.

ANGEWORK has published an additional verification article on the AI technology "TripoSplat," which generates 3D models from a single image. In the previous verification, challenges were observed in generating the backside, so this time we are retesting the success conditions and failure factors using AI-generated images. We analyzed the impact of factors such as single objects, shading, and viewpoint information on the generation results, and organized the conditions and points to note that make successful 3D conversion more likely. If you are interested in 3D generation from images, point cloud processing, digital twins, or spatial recognition AI, please take a look. ▼For more details, click here https://note.com/angework/n/n31a2cfaf87dd

Generative AI is rapidly evolving, with major services like ChatGPT and Claude frequently adding features and changing specifications. As a result, the methods of utilization that were optimal at the time of introduction quickly become outdated, leading to an increasing number of companies facing challenges in information gathering and operational improvement. "Re-improvement of Implemented AI," offered by ANGEWORK Inc., is a supportive service aimed at companies that have already implemented AI environments. It focuses on operational improvements, inspection and correction of AI-generated code, and the provision of analytical reports on AI-related news. Regarding the increasing AI-generated code, experienced technicians will check for vulnerabilities, design issues, and implementation mistakes that older generation AIs may have overlooked, addressing everything from cause investigation to corrections and operational verification. Additionally, important AI news and technological trends will be selected based on each company's industry and implementation status, and provided as weekly reports that include explanations of technical terms and risk analyses. The service not only helps companies implement AI but also supports continuous improvement and utilization. Initial consultations and sample reports are free of charge.

ANGEWORK has published a technical article on the support for 3D Gaussian Splatting (3DGS) in RealityKit, announced at WWDC26. Apple has added GaussianSplatResource to RealityKit, allowing 3DGS to be handled within the official framework. Additionally, collaboration with OpenUSD is progressing, and the potential for 3DGS to transition from research applications to general 3D workflows is beginning to emerge. The article explains the significance of RealityKit support, its relationship with OpenUSD, and the potential for standardization of 3DGS from our perspective as a company dealing with point cloud and LiDAR technology. ▼For more details, click here https://note.com/angework/n/nf984debd4a48

ANGEWORK has published a technical article on the paper "LGTM: Less Gaussians, Texture More," which was accepted at ICLR 2026. LGTM is a technology that achieves 4K quality representation while controlling the increasing number of Gaussians during high-resolution 3D Gaussian Splatting (3DGS). The article explains the fundamentals of 3DGS, the issues of increased load due to 4K conversion, the structure of LGTM, and the differences from conventional methods, all from our perspective as developers of the LiDAR and point cloud analysis system "mistrace." ▼For more details, click here https://note.com/angework/n/n9511604b4789

**Overview** ANGEWORK has published a technical article on the 3D reconstruction technology "GenRecon," announced by the Technical University of Munich and Huawei. GenRecon is a technology that generates 3D models with PBR materials from videos taken with smartphones. In comparison to traditional photogrammetry and LiDAR measurements, the article discusses: - Why high-quality 3D model generation is possible - How unseen areas are supplemented - The relationship with Trellis.2 - Points to consider for surveying and measurement applications This is explained from our perspective as developers of the LiDAR and point cloud analysis system "mistrace." If you are interested in 3D scanning, point cloud analysis, AI generation technology, or digital twins, please take a look. ▼ The article is here https://note.com/angework/n/ndab462656bbe