A conversation in. A structured, inspectable human model out — 11 dimensions, each carrying its own evidence, confidence and uncertainty.
Illustrative interface with animated sample data — not a live session. Every number KKOS publishes is measured and shown with its test conditions.
These are KKOS design targets. Each number is measured and published together with its test conditions — network, microphone and model version — once verified. Nothing here is a marketing estimate.
Target values, pending measurement: first-byte latency < 500 ms and single-session length 30 min. Verified numbers will replace these labels together with their test conditions.
Borrowed from how voice-AI quality is actually judged: reliability buys you in, expressiveness keeps the conversation flowing, and empathic understanding is what makes someone stay. KKOS is built on all three — and the top layer is inspectable.
Most voice AI measures how you feel; a few respond to it. KKOS shows you why it responded that way — and lets you overrule it.
lg3Voice, text and video in one flow — and a conversation you can interrupt, pause and resume like a real one.
lg2Eight doors, two levels: four product surfaces, four engineering deep-dives. Read exactly the level you came for.
Every dimension, the evidence behind it, the heat surface the engine reads, and the four-step pipeline.
Open the model → 🎙️The real voice catalogue behind the conversation — audition any of them before you pick one.
Browse voices → 🔐The six consents, what is stored, how keys are hashed, and what deleting a model actually clears.
Read the design → ✍️Voice, text or video — with a working text form that produces a real model on your account.
Start building → ⚙️The five engines that consume the model — capsule, temporal memory, causal discovery, dyadic attention and the correction loop — each with a claim you can test.
See the engines → 🏗️From one conversation to one match score in five layers, plus the master data chain that keeps every model, match and payment traceable.
See the architecture → 🎚️One core, three domains — and the correction loop that makes your disagreement part of the model instead of noise to be smoothed over.
See adaptation → 📈How a model changes: numbered revisions, the nine-year life-stage projection, the open research programme and an honest language roadmap.
See evolution →