Guides, news, and tutorials from AityTech — crafting code in Japan.
BM25 misses meaning. Vector search misses exact error strings. AgentKits Memory fuses both — but for Japanese, Chinese, and Korean queries, the keyword half of that fusion needs a second decision underneath it.
Dependency-mapping tools built around COBOL treat JCL as a thin wrapper — a job name and some DD statements. But condition-code branching, PROC nesting, and GDG generation windows carry real control flow of their own, and it's routinely left out of the graph. Why Legacy Dragon parses JCL as a first-class language in the same AST, not metadata bolted on afterward.
PrivateAI's text-to-speech tool covers 322+ voices across 142 languages entirely on-device. Here's why voice and language breadth — not raw speed — is the hard problem for browser-based TTS, and how that number stacks up against the open-source and paid alternatives.
The Marketing Kit's skills-registry.json catalogs 28 skills across five categories with a full dependency graph. Its own internal research doc names the reason for capping what gets loaded at once: 'context rot.' Here's how the shipped selector actually works, and why it's simpler than the research proposal that led to it.
iOS 26 and Google's Pixel Recorder now do on-device call transcription and summarization for free. Here's the specific boundary neither platform crosses — and why that's exactly where MinuteAI's Chrome Extension operates.
METI's 2025 digital cliff warning gets read as workforce attrition and rip-and-replace math. The failure mode that actually corrupts migrations first is smaller and easier to miss: EBCDIC and Shift-JIS don't even agree on whether letters or numbers sort first. Why Legacy Dragon treats character encoding as a parsing-layer concern, not a pre-processing step bolted on later.
Cloud AI is priced by the token because every query costs the vendor real compute. On-device tools don't have that bill. Here's what that structural difference actually buys — and doesn't — for a product like PrivateAI.
Open the agentkits-marketing repo's plugins/ folder and you'll find two self-contained products — Content Factory and Campaign Manager, each with its own commands, agents, and roadmap. Open marketplace.json and there's exactly one installable plugin. Here's what that gap says about where the kit is headed, and why it fits a wider shift away from monolithic tool bundles.
AgentKits shipped its Marketing Kit first — 18 agents, 93 commands, 28 skills, tagged Ready. The Engineer Kit sits right next to it on the homepage with a bigger skill count (40) and a yellow Preparing badge. Why build the coding-agent kit second, into a market that's already full of coding agents?
The diarization feature is one story. Building, validating, and releasing the 743MB Python/PyTorch runtime behind it is another. Here's the pipeline MinuteAI runs before that download ever reaches a user's Mac.
RBS's 2012 meltdown and TSB's 2018 migration disaster didn't fail because the underlying COBOL was too old to touch. They failed because nobody could see the full blast radius of a change before it shipped. Why Legacy Dragon's AST graph is built as impact analysis, not just a translation aid.
Since Chrome 148, Summarizer, Translator, and Language Detector are stable built-in browser APIs running on Gemini Nano — no server required. Here's what that platform-level shift actually covers, and where a broader tool suite like PrivateAI still has to fill the gap.
AgentKits ships its README and 23-module training course in ten languages — not through a translation vendor, but through a Node.js tool built on the Claude Agent SDK that the team also open-sourced on its own. Here's how the pipeline actually works, and why an LLM is the right tool for this specific job.
Every major mainframe vendor now sells an LLM that rewrites COBOL into Java for you. The pitch is fast migrations — but generative rewrites aren't reproducible, and reproducibility is exactly what a billion-dollar batch job can't do without. Why Legacy Dragon starts from a deterministic AST graph instead.
Most privacy-focused meeting recorders still send the transcript to a cloud LLM to generate the summary. MinuteAI's GGUF/llama.cpp pipeline keeps that step on-device too — here's why that gap matters more than it sounds.
ChatGPT alone triggered over 410 million DLP policy violations in 2025. Most of that leakage isn't malicious — it's employees doing normal work through tools that happen to have a server in the loop. Here's why PrivateAI's browser-only architecture sidesteps the problem structurally.
Mainframe COBOL gets the modernization spotlight, but most organizations running it also run a second, quieter legacy estate — Windows-era VB6 and PowerBuilder apps. Why Legacy Dragon parses both categories with the same AST-based approach.
AgentKits' installer supports 18 different AI coding tools from one canonical .claude/ source — generating native launcher files for most, and an actual AGENTS.md for the two that read it. Here's how the config-driven system works, and how it relates to the AGENTS.md standard.
MinuteAI's speaker diarization runs on pyannote.audio and PyTorch — too heavy to bundle into a lightweight SwiftUI app. Here's why it ships as an optional 234MB download instead, and how that fits a wider shift toward on-device diarization.
COBOL systems don't have an expiration date, but the engineers who know why they work that way do. Why Legacy Dragon's AST graph matters as much for capturing retiring expertise as for migration speed.
AI agent memory has become a crowded 2026 category — Mem0, Zep, Letta, Cognee, LangMem — spanning managed cloud, self-hosted, and local MCP. Here's why AgentKits Memory stays in the local-MCP lane, and what that trade-off actually costs and buys.
WebGPU finally makes browser-based AI fast, and a wave of 2026 data-localization law makes it valuable. Why on-device tools like PrivateAI are arriving at the right moment.
MinuteAI's Chrome Extension captures your own audio and streams it to localhost — no bot joins the call. Here's the architecture, and why 2026's platform crackdown on meeting bots makes that design look prescient.
The shops still running COBOL and JCL are also the shops with the tightest network controls in software. A parser that needs npm install or a cloud API call is dead on arrival there — which is why Legacy Dragon isn't built that way.
Most open-source Claude Code agent repos group agents by technical domain. AgentKits' Marketing Kit groups them by TOFU/MOFU/BOFU instead — a deliberate bet on depth over breadth.
How MinuteAI turns meetings, interviews, and voice memos into text entirely on-device — the model pipeline, the memory constraints, and why none of it touches a server.
Zunavi's feature flags don't just turn features on and off — they draw an explicit line between capabilities that are reliable today and ones that are still assistive and require human verification. Here's why that distinction is enforced twice.
Most drawing-management SaaS tools assume your files live in someone else's cloud. For semiconductor and precision manufacturers, that assumption is a non-starter. Here's how Zunavi's architecture makes on-premise the default, not an afterthought.
The legacy system behind Zunavi tracked the 'latest' drawing revision by deleting and reinserting rows. Here's why that pattern breaks, and how PostgreSQL, Prisma, and a single boolean flag replace it for good.
Why parsing is the bottleneck in legacy modernization projects, and how AST-based tools built for COBOL, JCL, PL/I, and other mainframe-era languages change the economics.
A look at the real git history, commit cadence, and automation behind AityTech's product portfolio — conventional commits, an agent-driven workflow, and where AI-assisted tooling actually pulls its weight.
A closer look at PrivateAI's browser-based OCR, text-to-speech, speech-to-text, background removal, and translation tools — and why none of it touches a server.
PrivateAI, Legacy Dragon, AgentKits, Yomite, and StudyKits — a look at the AityTech product lineup and the privacy-first, on-device philosophy that connects them.
Zunavi (図ナビ) is our in-development, on-premise-first platform for managing engineering drawings, specs, and approvals in manufacturing. Here's the problem it's built to solve.
How we architect macOS applications that process AI workloads entirely on-device — no cloud uploads, no data collection, no compromise on capability.
Introducing our blog — a space for guides, news, and tutorials on iOS development, web platforms, AI solutions, and open-source tools from our studio in Japan.