What we build, and what we operate.

Waguri is our principal product. It is an autonomous moderation system that reviews what is posted in a community, decides whether it breaches that community's rules, and acts on the result, while remaining accountable to the humans who run the server. It is presently deployed across active communities and processes several million messages.

Content moderation

Every image, video, and animated attachment posted in an enabled channel is examined. A fast local classifier performs the first pass. Anything it flags is escalated to a stronger vision model, which produces an independent judgement and a written description of what it sees. A separate reviewing model then reads that description and decides whether the content genuinely breaches the rule in question.

The order of those steps is deliberate. No punitive action is taken until the reviewing stage confirms the finding. A message is not removed, and a member is not muted, on the strength of a numeric score alone. Where the reviewer judges a flag to be a false positive, the message is left untouched, no penalty is applied, and the item is recorded so the same content is not re-examined. This ordering exists because an incorrect automated punishment is more damaging to a community than a delayed one.

Text is assessed in parallel for prohibited language, threats, and coordinated abuse. Text rendered inside images is extracted and assessed by the same rules, so material posted as a picture is not treated differently from material posted as a sentence.

Child-safety controls

Known child sexual abuse material is matched by cryptographic hash against established industry hash sets before any other processing occurs, and the check is performed against the hosting platform's own content address rather than by retrieving the file. Amanaric does not retain user media. Images are not written to our storage and are not re-uploaded by our systems; our services reference the originating platform's content and hold no copy. A positive match is contained automatically and is not displayed to moderators. Where a report to the National Center for Missing and Exploited Children is required, it is made by a person, not by an automated process.

Safety and welfare handling

Material indicating self-harm or a member in crisis is routed away from the punitive path entirely. Such a post is removed from public view so that it is not amplified, but no penalty is issued and no automated warning is sent. The matter is instead referred privately to a human moderator with the context needed to respond appropriately. A person in distress is not treated as a rule-breaker.

Accountability and appeals

Every automated action is recorded with the reasoning that produced it, the evidence considered, and the model verdicts involved. Moderators can review that record and reverse any decision. Members may appeal, and appeals are re-examined against the original evidence rather than dismissed. The ability to correct a mistake is treated as a requirement of the system, not an optional feature. Appeals may also be submitted through our support desk.

Conversation and voice

Waguri holds ordinary conversation in text and in voice channels. Voice operates as a full spoken exchange: speech is transcribed, a reply is composed, and the reply is spoken aloud. Transcription runs primarily on our own hardware, with a commercial provider as an automatic fallback when that hardware is unavailable. Audio is processed transiently for transcription and is not retained as audio.

Consent and data handling

Conversational features are gated on acceptance of the Terms of Service. Members who have not accepted them still receive replies, in a restricted and deliberately unobjectionable mode, and no personal information is retained for them. Actions performed on a member's behalf likewise require acceptance; where a request cannot be carried out for this reason, Waguri states so plainly rather than implying the action was completed. Moderation is never gated in this way: protection of a community does not depend on the consent of the person breaching its rules.

Administration

Server administrators configure Waguri through a web dashboard: which channels are examined, the sensitivity of each control, the persona and language used in conversation, and the escalation path for serious matters. Configuration is held per community; no setting is shared between servers.

Anime Hideout is a web community platform with member accounts, profiles, direct messaging, voice and video calling, moderation tooling, and a formal appeals process. It is accompanied by a desktop application distributed with its own installer and automatic update channel.

The platform also provides the account system used across our products. Authentication is performed once against Anime Hideout, and other properties obtain a session through a signed single-use token rather than holding separate credentials. This keeps one authoritative account record per person and avoids duplicating personal data across sites.

Kurono is an operating system written from scratch in freestanding C++17 and x86 assembly. There is no C standard library in the kernel, no borrowed POSIX runtime beneath it, and no existing kernel reused as its core. It combines a native desktop operating system, a Linux-compatible syscall and runtime layer, a Windows command environment, a package and update pipeline, and a Type 1 hypervisor into a single bootable image.

Booting and hardware

Kurono boots on real x86_64 hardware as a Multiboot2 ELF kernel, and additionally produces standalone EFI loaders, an emergency recovery boot path, and installer targets that deploy onto FAT32 and ext4 volumes. Virtualisation under QEMU with KVM is the day-to-day development and continuous-integration environment; physical hardware is the target.

Nineteen core hardware drivers are implemented, together with a wireless network interface layer, a complete TCP/IP stack, and a multi-backend display manager supporting BGA, VirtIO GPU, Intel, NVIDIA, and AMD devices. Hybrid graphics topologies, including NVIDIA Optimus and AMD PowerXpress, are detected and handled.

Symmetric multiprocessing

All processors boot and execute kernel code in parallel, with genuine application-processor bring-up by the INIT-SIPI-SIPI sequence, a per-processor SYSCALL entry path, per-processor descriptor and task-state structures, and a per-processor current task. Secondary processors run ring 3 Linux user processes concurrently with the desktop, and are preemptive: each arms its own local timer to time-slice the threads it runs rather than switching only at system calls.

Running existing software

A Linux syscall runtime provides concurrent threads and futexes, event polling, write-exclusive-or-execute memory protection, and file-descriptor passing between processes. An ELF64 dynamic linker is implemented inside the kernel. A Wayland compositor, likewise in-kernel, renders genuine Wayland clients compiled against musl, with the PulseAudio and D-Bus runtime services those clients expect.

The most demanding demonstration of that stack is Firefox: a full Firefox 140.11 ESR build, compiled against musl and Wayland with software WebRender, boots across the secondary processors, loads a real profile, retrieves a page over the in-kernel TCP/IP stack, and paints that page through the in-kernel compositor. Reaching it required three kernel-level corrections, each exposing a genuine defect: a frame-handoff barrier for threads migrating between processors, correct buffer-release signalling so the client does not exhaust its buffer pool, and per-surface shadow buffers so a damage-incremental client cannot blank its own window.

Alongside this, more than one hundred and fifty shell commands are implemented across the native Kurono, Linux, and Windows command environments, with an Alpine guest path and an on-demand Debian root filesystem and update pipeline.

Security work

Kurono has undergone a whole-system security audit. Approximately sixty-eight issues were triaged and around twenty confirmed memory-safety defects have been corrected, including integer-overflow heap overflows in the media codecs, two remotely triggerable stack overflows in the D-Bus implementation, a guest-to-host escape through the 9p filesystem, graphics base-address-register mapping faults, unsafe trust of on-disk structures in the ext4, KFS, and NVMe paths, recursion and overflow defects in the KCL and KJ interpreters, and a stack-to-heap correction in the shell. That work is ongoing.

Present status

We state this plainly: Kurono is not production-ready, is not free of defects, and is not a daily-use operating system. It is a long-running single-developer project presently in a stabilisation phase, in which subsystems built up over time are being debugged and released as each becomes sound. Known limitations remain, including intermittent slowness in boot and an unresolved stall during multiprocessor bring-up. It is presented here as an ambitious work in progress in which the difficult parts largely exist and the remaining effort is finishing and correction.

Kaiten introduces NVIDIA DLSS 4, comprising both Super Resolution and Frame Generation, to Minecraft Java Edition 1.21. It is built upon a fork of VulkanMod together with the NVIDIA Streamline software development kit.

The work consists of integrating a contemporary upscaling and frame-generation pipeline into a rendering path that was never designed to accommodate one, which requires operating at the level of the graphics API rather than the game.

We operate our own mail platform rather than renting one. It comprises a mail server we administer, a webmail client for which we wrote the mail-protocol driver ourselves, and a multi-tenant control panel for managing domains and addresses.

Sender authentication is configured to enforcement across every domain we operate, with reverse DNS aligned and message signing verified by external round-trip testing rather than assumed. Ownership of a domain is verified through the domain name system before any address on it is provisioned.

We own and operate the infrastructure beneath our products: application servers, database systems, and graphics processing hardware for machine-learning inference, with commercial inference capacity used as an overflow tier. This allows us to control cost, latency, and the handling of data, and it is the reason we are able to state plainly where any piece of information resides.

Public traffic terminates at a single reverse proxy with a content delivery network before it. Internal services bind only to the local interface and are not otherwise reachable. Credentials are held in restricted environment files and never in source control.

Application hosting for external customers is not presently offered. Enquiries regarding infrastructure work may be directed to [email protected].

We accept a limited number of commissioned engagements where the work is a good fit for our expertise. Each begins with a written scope agreed by both parties, which defines the deliverables, the timeline, and the terms before any work commences.

Enquiries should be sent to [email protected] setting out the objective, the required functionality, and the intended timeline. The terms applying to commissioned work are set out in the Custom Project Agreement.