.svg)
003JAPANESE MANUFACTURING COMPANY · 2025
800+
JAPANESE MEETINGS ANALYZED
AI Meeting Dashboard with Kanji Verification and Action-Item Kanban
RAW TRANSCRIPT → AI AGENT → KANJI-VERIFIED KNOWLEDGE
9("&§$/()"&$§)"§!?@#%^*{}<>|~±²³µ¶·¹º»¼½¾¿×÷&§$?!("&§$/()"&$§)"§!?@#%^*{}<>|~±²³µ¶·¹º»¼½&§9("&§$/()"&$§)"§!?@#%*<>|~$§&!?()9("&§$/()"&$§)"§!?@#%^*{}<>|~±²³µ¶·¹º»¼½¾¿×÷&§$?!("&§$/()"&$§)"§!?@#%^*{}<>|~±²³µ¶·¹º»¼½&§9("&§$/()"&$§)"§!?@#%*<>|~$§&!?()
AI AGENT
皆様、本日はお忙しい中お集まりいただきありがとうございます。本日の議題は品質管理の最新状況についてです。まず、先月の生産実績をご報告いたします。タイ工場の稼働率は前月比で改善しております。皆様、本日はお忙しい中お集まりいただきありがとうございます。本日の議題は品質管理の最新状況についてです。まず、先月の生産実績をご報告いたします。タイ工場の稼働率は前月比で改善しております。
Kuroko Labs built a fully automated meeting dashboard for a Japanese manufacturing company on top of self-hosted Whisper: Teams records every meeting, a cron job fetches the recording via Microsoft Graph, and local Whisper transcribes it with Japanese-optimized accuracy for a five-stage agent pipeline. That pipeline cleans up raw transcripts, verifies kanji names against a directory of 73 employees, extracts tasks into a Kanban board with effort estimates and writes content into a company-wide knowledge agent. Microsoft Graph Attendance controls who is allowed to see which meeting, in Japanese, Thai and English. From 800+ meetings, this produces searchable, accurate, actionable knowledge.
800+
Meetings analyzed
Fully automated since 2025
5
AI agents
Transcription, cleanup, kanji, synthesis, action items
8
Meeting types
Automatically classified & color-coded
73
Employees in directory
Kanji ground truth for verification
3
Languages in the dashboard
Japanese · Thai · English
01 — THE CHALLENGE
800 Japanese meetings, one CEO who cannot be everywhere
The client runs more than 80 meetings a month via Microsoft Teams, between Japan and Thailand, almost entirely in Japanese. Client appointments, production planning, quality checks, supplier coordination. The CEO needs to know what was discussed, but it is impossible for him to attend them all.
Automatic transcription did produce text, but it was riddled with errors. Japanese names were systematically rendered into the wrong kanji: similar-sounding characters were confused, and correct employee names were warped into unreadable strings. Anyone who did not already have the context in their head could not decode the transcript.
The real problem ran deeper: meeting knowledge evaporated. Action items ended up in Slack, decisions in private notes, context disappeared entirely. No central repository, no per-meeting searchability, no reliable summaries, and no control over who was allowed to see which sensitive content.
The client needed a system that analyzes Japanese meetings fully automatically, verifies kanji, extracts tasks and makes everything searchable, with correct names, targeted follow-up questions, a task board and access control derived from actual attendance.
800+
Meetings out of reach
The CEO cannot be everywhere, yet has to know everything
漢
Wrong kanji recognition
Speech recognition confuses names like 旭化成 and 朝日課税
?
Action items lost
Who does what by when? Scattered across Slack, Teams chat, email
∅
No company memory
Meeting content evaporates, nothing is searchable later
02 — THE SOLUTION
Five agents from the audio track to the task
Kuroko Labs built a five-stage agent pipeline that processes every meeting just minutes after it ends, from the audio track all the way to a searchable knowledge base, with a task board, audio jump markers and kanji-verified names.
Step one: a cron job checks the company calendar every five minutes and automatically invites the Whisper (self-hosted) bot into every upcoming Teams meeting via Microsoft Graph, in Japan just as much as at the two Thai sites. Whisper (self-hosted), a meeting-transcription service specialized in Japanese, records and returns the finished transcript through a secured webhook: token authentication, replay protection against duplicates. A three-stage matching engine assigns every transcript to the right calendar entry, first via the join URL, then via time window and organizer, finally via title similarity. The first agent then cleans up the raw transcript: filler words (えーと, あの, はい) removed, speakers grouped, sentences normalized.
Step two is the kanji verification agent, and the reason the whole project works at all. It receives the cleaned-up transcript and runs a deep analysis against an employee directory of 73 entries. Every recognized name is checked against ground-truth kanji and corrected deterministically: 旭化成 (Asahi Kasei) instead of 朝日課税, 西端卓也 (リテール企画販売部) instead of 西畑拓也. Three lines of defense interlock, the translator system prompt, the unified prompt in the action extractor and a final
correctNames() pass.Step three: Claude Sonnet 4.6 produces a contextualized summary using textile industry terminology. Terms like “thread count”, “MOQ” or “AQL level” are carried correctly into the Japanese summary, no clumsy pseudo-translations. For meetings longer than three hours, the system uses chunked summarization: logical sections are summarized individually, then consolidated.
03 — ARCHITECTURE
Meetings become work, the task Kanban
01AUTO-RECORDING
Teams recordings, transcribed on-premise
A cron job fetches the Teams cloud recordings via Microsoft Graph, including at the Thai sites, and hands them to self-hosted Whisper, which transcribes with Japanese-optimized accuracy. There is no public endpoint accepting transcripts: the pipeline pulls. A three-stage matching engine assigns every transcript to the right calendar entry. The first agent then cleans up the raw transcript: fillers (えーと, あの, はい) removed, speakers grouped, sentences normalized.
800+
Meetings captured
02KANJI AGENT
Verification against the employee directory
A dedicated agent checks every Japanese name against an employee directory of 73 entries and corrects systematic errors deterministically, e.g. 旭化成 (Asahi Kasei) instead of the mistaken 朝日課税, or みずほ銀行 (Mizuho) instead of 水穂銀行. Three lines of defense work together: the translator system prompt, the unified prompt in the action extractor and a final
correctNames() pass.73
Employees in directory
03EFFORT KANBAN
Meetings become work, automatically
Claude Haiku 4.5 extracts up to 30 tasks per meeting in a single call: title, description, estimated effort 1–5 (from 軽微, 15–30 min, up to 大型, 1+ week), complexity, blocker risk and due date. Assignment happens via fuzzy matching against the attendee lists, no manual mapping needed. Three display modes: full, compact, list.
5
Effort levels
04PER-MEETING Q&A
Fred, the meeting assistant
Using a unique meeting ID, the system filters precisely to a single recording. “What was decided about supplier X?”, the agent answers with a timestamp and an audio jump marker straight to the relevant passage. Follow-up questions continue in the same thread, with rate limiting, a circuit breaker and a fallback to English sources when the Japanese track is thin.
100%
Meeting-isolated answers
05ACCESS CONTROL
Only your own meetings, or granted access
In the calendar, every employee sees only the meetings they attended. Visibility is derived from Microsoft Graph attendance records plus speaker detection in the transcript, even silent participants (マイク無音) are recognized as present. The CEO and department heads receive targeted whitelisting for strategically relevant meetings.
0
Unauthorized accesses
06DUAL RAG
Company memory + per-meeting index
Every kanji-verified transcript flows into the company-wide RAG on Supabase, vector search plus Japanese full-text search, sharpened by Cohere reranking. Four answer tiers from Flash to Deep balance speed and depth, and Redis caching delivers recurring answers in milliseconds. In parallel, every meeting gets its own RAG index, and Outlook mailboxes also feed into the company memory after a two-stage relevance filter.
2
RAG systems in parallel
Step four is the most exciting: the action-item extractor. Instead of classic prompt chaining, it uses a single-pass call with Claude Haiku 4.5 that produces tasks, descriptions, effort level, complexity, blocker risk and due date in one step, kanji-safe, deterministic, with assignee resolution via fuzzy matching against the attendee list.
The effort levels 1–5 are the heart of it: 1 軽微 (15–30 min, an email, a confirmation), 2 小規模 (1–3 hours), 3 標準 (half a day to a full day, a standard review, multi-person coordination), 4 重作業 (2–3 days, material procurement, multi-department coordination), 5 大型 (1+ week, a new project, a major build). Up to 30 tasks are extracted per meeting, quality over quantity.
The Kanban board has four states, やること (to-do), 進行中 (in progress), 確認待ち (review), 完了 (done), and three display modes: Full (cards with all details), Compact (condensed cards for 30+ visible tasks) and List (tabular, sortable by effort and due date). A translation layer based on GPT-4o mini translates Kanban content between Japanese, Thai and English, the entire dashboard is localized in all three languages. Completed cards are archived automatically and quietly deleted later.
LIVE PREVIEW
Calendar Dashboard — Meeting Overview
MonTueWedThuFri
● Click on a day to see the details
LIVE PREVIEW · ACTION-ITEM KANBAN
From meeting to work — across four lanes
The action-item agent reads every meeting from the calendar above and extracts tasks with effort level, due date and assignee. The half-year review becomes the Q2 cost-sheet task, the infrastructure check becomes the cloud-monitoring ticket. This turns appointments into actionable cards automatically — in the right lane, with the right person.
04 — RESULTS
Calendar dashboard with access control
800+
Meetings processed fully automatically
3
Tenants: Japan + two sites in Thailand
5
Effort levels in the task Kanban
0
Manual post-processing required
METRICBEFOREAFTER
Meeting access for the CEOOnly when attending in personVirtually present in every meeting
Kanji accuracySpeech recognition garbles namesVerified against a 73-employee directory
Action itemsScattered across Slack, email, notesKanban with effort levels 1–5
Knowledge archivingContent gets lostDual RAG: company-wide + per meeting
Follow-up questionsSearch the transcript manuallyQ&A with timestamp & audio jump
The dashboard presents all meetings in a Japanese calendar view with color-coded meeting types. The eight automatic classifications provide structure: 営業 (sales), 戦略 (strategy), 製品 (product), 財務 (finance), 人事 (HR), 技術 (engineering), マーケティング (marketing), その他 (other).
The decisive piece is the attendance-based access control: every employee sees only the meetings they actually attended. Visibility is governed automatically via Microsoft Graph attendance records and speaker detection in the transcript. Even silent participants (マイク無音), people who joined with their microphone off, are recognized as present.
For the CEO and department heads there is whitelisting: they can be granted access to strategically relevant meetings without having attended them. Self-registration with Microsoft Graph OTP verification ensures that only authorized employees with a company email gain access, in Japan as in Thailand, with no manual admin effort. Before every upcoming meeting, the system also automatically generates a meeting briefing with context from earlier meetings, precomputed in the background so the dashboard card loads without any wait. A PM2 cluster with Redis-backed rate limiting keeps the system stable even during sync spikes.
“
I can now be in every meeting without physically attending. I ask a question, get the answer with a timestamp and can jump straight into the audio recording. That has fundamentally changed the way I run our company.CEO, JAPANESE MANUFACTURING COMPANY
TECH STACK
Whisper (self-hosted)
Japanese meeting transcription, on-premise
Claude Sonnet 4.6
Synthesis & kanji verification
Claude Haiku 4.5
Action items & RAG answers
GPT-4o mini
Query planner, translation & fallback
Microsoft Graph
Calendar, attendance & OTP auth
Supabase + pgvector
Dual RAG index & full-text search
Cohere Rerank
More precise RAG hits
PM2 Cluster
Multi-worker backend
Redis
Caching & rate limiting
SQLite (WAL)
Meeting metadata & Kanban state
Node.js + Express
Backend & SSE streaming
OUTLOOK
The platform keeps growing: SharePoint and Drive documents are next in line to feed the company-wide RAG, so that meetings, emails and documents become searchable in a single company memory. Also planned is cross-meeting analysis that tracks topics across multiple sessions, such as the escalation curve of a supplier discussion over four weeks, along with a cost-analytics layer that makes LLM consumption transparent per request. In the long term, the dual RAG becomes the central memory of the entire company group: every meeting, every decision, every correctly spelled name, available anytime, in three languages.
NEXT STEP
A similar project?
GET IN TOUCH ALL CASE STUDIESFrequently Asked Questions — AI Meeting Dashboard Case Study
How does the kanji verification work?+
A dedicated AI agent checks every cleaned-up transcript against an employee directory of 73 staff members. Japanese names that the speech recognition systematically mistranscribes are corrected deterministically — for example „西畑拓也“ to „西端卓也“. This ensures that every transcript contains the correct kanji names.
Who sees which meetings in the dashboard?+
In the calendar dashboard, every employee sees only the meetings they actually attended. Visibility is controlled automatically via Microsoft Graph attendance records and speaker recognition. Through whitelisting, the CEO or department heads can be granted access to additional meetings.
How is meeting data used in the RAG?+
Every kanji-verified meeting is fed into a company-wide RAG agent on Supabase — as central company knowledge. In parallel, each meeting gets its own dedicated RAG by meeting ID. This makes both cross-meeting searches and targeted per-meeting queries with timestamps and audio links possible. Outlook emails also flow into the company memory after a two-stage relevance filter.
Can Kuroko Labs build a similar meeting dashboard for my company?+
Yes — the architecture of AI analysis pipeline, kanji verification and attendee-based access control can be applied across any industry. Whether for executives who cannot attend every meeting, or for teams with international locations: Kuroko Labs builds a tailor-made solution. The initial consultation and potential analysis are free of charge.
How secure is access to the meeting dashboard?+
Self-registration with Microsoft Graph OTP verification ensures that only authorized employees with a company email gain access. Transcripts are pulled via Microsoft Graph rather than accepted through a public webhook, there is no inbound endpoint; attendee-based visibility, PM2 cluster mode and Redis-backed rate limiting guarantee security and stability.
Technology used
Processing uses Whisper self-hosted, Microsoft Graph for the Microsoft 365 connection and Supabase. Speech recognition runs locally; the recordings never leave the building.
The links lead to the documentation of each component.