In Japan, resignation agencies are now a thing — employees pay third parties to quit for them. To me, that's the visible symptom of communication breakdown companies notice too late.

Six years ago I watched it from a board seat — the best people left quietly while every dashboard said fine. So I'm building Kashi: we analyze the structure of how teams talk — turn-taking, who gets interrupted, who stops speaking — never content, never emotion.

Our moat is three architectural properties LLMs structurally cannot match: auditable, reproducible, content-blind. The doctrine is baked into our data layer, not into a prompt.

Pre-pilot, 908 tests, 25% reply rate on customer discovery. Filipino-Japanese, Keio business undergrad, building with a co-founder-level teammate.

Kashi helps companies see warning signs before silence becomes resignation.

Hi, I'm Justine, and I'm building Kashi.

In Japan, we now have something called resignation agencies — where employees pay a third party to quit their job on their behalf. I see it as the visible symptom of a deeper dialogue breakdown inside organizations. Companies communicate more than ever, but cannot see when communication is starting to break down. By the time HR or executives notice, the best person has already mentally left.

I'm building Kashi because I lived this. Six years ago I joined the board of a small company in Japan and watched our best people leave, quietly, while every dashboard said fine. I'm Filipino-Japanese, finishing Keio business school, building with a co-founder-level teammate. No one else I know sits in both cultures and has spent six years staring at this exact gap.

Kashi analyzes the structural patterns in synchronous meetings — turn-taking, who gets interrupted, who stops speaking. Never content. Never emotion. Never employee scoring.

Why won't GPT eat us? Three architectural reasons.

One — our moat is three properties LLMs structurally cannot match: auditable, reproducible, content-blind. LLMs are probabilistic by design — same input, different output — opaque, and read content by design. For regulated workplace use, those aren't feature gaps — they're regulatory deal-breakers.

Two — longitudinal baselines require persistent state, role-bounded storage, k-anonymity, audit logs. By the time a GPT wrapper builds equivalent state, we have N years of baseline they don't have.

Three — our doctrine is baked into the data layer: SQL + RLS + contracts physically exclude individual scoring, ranking, HR decisions. A GPT prompt can be rewritten next month. Ours can't. Moat is what we refuse to do, and the fact that we can prove we can't do it, not just that we don't.

Our regulatory lane is EU AI Act Annex III §4 — workplace AI monitoring as high-risk. We were designed to comply from day one.

Our outcomes are concrete: in the named ICP — Tokyo knowledge-work SMBs of 50 to 500 — pilots target Signal-to-Action within SLA above 80%, voluntary turnover reduction of 10 to 20 percent, employee NPS above 7.

Success has three phases: 5 paid pilots in 12 months; 50 customers in 24 to 36 months, 30 via 産業医 / 社労士 / coach reseller channel; in 5 years, the default trusted lane when regulation lands.

Pre-pilot — 908 tests, 25% reply rate on customer discovery.

Kashi helps organizations see the warning signs before silence becomes resignation. Thank you.