Structural pattern detection for workplace teams.
Enterprise B2B · Japan-first · designed for global compliance.
Employees pay third parties to quit their job on their behalf.
At first it sounds like a strange trend.
It's the visible symptom of dialogue breakdown inside organizations — by the time HR notices, the best person has already mentally left.
Companies communicate more than ever — more meetings, more Slack, more tools. But they cannot see when communication is starting to break down.
People stop speaking up.
One person dominates.
Same issue recurs unresolved.
Friction nobody can name.
In Japan: hierarchy + generational gaps. In the U.S.: cultural + communication norms. Same structural pattern.
Meeting transcripts + metadata. Used once.
Turn-taking · response latency · floor-share asymmetry · interruption direction.
30 / 90 / 180-day patterns surface where a single meeting looks like noise.
Never judges people. Never predicts who will quit. Never diagnoses emotions. Detects structural antecedents of dialogue breakdown.
Demo tenant live in production. Stress-tested at 50 meetings / 5,000 turns. 908 reproducible tests.
| What we see | What we don't | Why |
|---|---|---|
| Team-level structural change | Meeting content · individual utterances | No Payload — not stored |
| Longitudinal pattern shift | Emotion · mood · psychological state | EU AI Act Annex III §4 high-risk scope |
| Role-scoped reports | Individual score · HR decision | SQL + RLS + contracts exclude by design |
The limits aren't a UI setting. They live in the data layer. Cannot be lifted by an admin toggle.
LLMs are probabilistic, opaque, content-reading by design. For regulated workplace use, those aren't feature gaps — they're regulatory deal-breakers.
Persistent state + role-bounded storage + k-anonymity + audit logs. A weekend wrapper can store state; it cannot reproduce validated, customer-specific baselines tied to documented manager actions.
SQL + RLS + contracts physically exclude individual scoring, ranking, HR decisions. A prompt can be rewritten next month. Ours can't be — without rebuilding the product.
The real moat compounds OVER architecture: validated longitudinal baselines × renewal-proven trust × reseller distribution. We don't pretend to have it pre-pilot — we're building toward it.
Run rule-based feature extraction on transcripts.
Preserve inputs + versions.
Generate an audit trail.
Mimic the role-bounded views.
Even add tenant isolation + RLS.
Microsoft can do all this beside its Viva product, using meeting telemetry it already owns.
Validated longitudinal baselines — customer-specific, tied to documented management actions and renewal outcomes.
Governance evidence proving the architecture wins procurement.
Reseller-defended distribution through 産業医 · 社労士 · org-dev coaches.
Honest framing of the bet: employers and workers will not trust the meeting-platform owner to govern the diagnostic layer.
Microsoft can copy the analytics. We're betting they cannot copy the trust. That bet is proven by renewals — not by architecture claims.
Rising turnover cost.
Taishoku-daiko (resignation-agency) services — itself a visibility signal.
MHLW Power Harassment Prevention Law expanded to SMB in 2022. ~3.67M Japanese firms newly liable.
EU AI Act Annex III §4 classifies workplace AI monitoring as high-risk. Kashi was designed to comply from day one.
Indirect-communication culture amplifies the gap. "Records exist, but evaluation-AI is scary" — Kashi's clean-boundary design fits exactly in that gap.
Phase 1 is gated on renewal — not "5 pilots". That renewal proves signal-to-action validity. Without it, the rest is premature.
No paying customers yet. Next 12 months are about converting these conversations into 5 paid pilots — gated on Phase 1 renewal milestone.
Before silence becomes resignation.
Thank you.