Yunjue Technology · 云玦科技

Your Life, Rendered as Context

Yunjue is building a personal agent that grows with you — always-on hardware on your body, living in your phone, evolving in the cloud. Not another AI assistant, but a new kind of personal AI infrastructure: a long-term companion that, given enough time, actually understands who you are.

What we build

An agent that grows with you, not a mirror you stare at

Mainstream wearables are mirrors: they reflect heart rate, steps, and sleep numbers — but never tell you why, and never grow with you. Yunjue works in two layers: the always-on hardware is an external sensory system — continuously watching, listening, and measuring without interrupting your conscious flow; the cloud AI is an external prefrontal cortex — turning raw signals into reflection, review, and personalized decision support, so the agent actually understands you and grows with you.

Early
Multimodal depth
Raw-signal fusion, not Late Fusion
Zero
Personalization
Zero-Skill start, fully personal in 14 days
Nightly
Self-evolution
Nightly self-evolution sandbox · audit events · update profile · author tools and skills

Three-layer moat

From signal to application, layer by layer

01 / Signal

Early Fusion multimodal

Not text-then-concat

Mainstream Late Fusion compresses audio, heart rate, vision into text first, then concats — losing all temporal and causal links. Yunjue aligns 7 modalities at the raw-signal layer, so conclusions can be event-, moment-, and person-precise.
02 / System

Self-Evolving Agent

In-Situ Self-Evolving · 5 benchmarks, 3 SOTA / 2 runner-up

Our research work, Yunjue Agent (In-Situ Self-Evolving), proposes "inference IS evolution" — measured against GPT-5.2 Pro, Gemini 3 Pro, and other frontier baselines across HLE / DSQA / FSC / xSciQA / xDS; paper, code, and full evolution traces are open source. In production, a dedicated self-evolution sandbox runs every night for each user: audits today's events, incrementally refreshes the user model, identifies capability gaps, and authors tools and Skills on the spot.
03 / Application

Zero-Skill personalization

Each card is a standalone HTML

Day-one feed is empty. Days 3–7 the system starts building Skills. By day 14, no two users share the same card library. Tool users see review cards; companion-style users see emotional mirroring; creators see story prompts.

Each layer depends on the one below. Without Early Fusion the agent sees no dense signal; without self-evolution the application stays one-size-fits-all. Full reasoning on the tech page.

From Early Fusion to Human-Centric World Model

What we can do today, where we're headed

Today · Early Fusion

Mainstream multimodal AI uses Late Fusion: audio, heart rate, vision are each compressed to text first, then concatenated — losing all temporal and intensity details. Yunjue aligns 7 modalities at the raw-signal layer, making conclusions precise to event, moment, and individual.

Self-report vs. body

You say "felt fine," but the heart-rate curve says stressed. What you report in words and what your body actually shows are aligned in the same window — the gap language never catches is surfaced.

Same activity, across days

The same sit-and-type session, today vs. the 14-day baseline, can look very different physiologically. "Today vs. normal" becomes a measurable quantity.

Multimodal causal chains

"Heart rate rose 3 seconds before that sentence was spoken" — heart rate, voiceprint, and vision aligned at the raw layer construct causal chains language alone never produces.

Long-term goal · Human-Centric World Model

Today we borrow the alignment capabilities of mainstream multimodal models to run Early Fusion while accumulating dense "unusual modalities → human behavior outcome" labeled data. When the data flywheel matures, Yunjue's long-term goal is to train a truly human-centric multimodal foundation model — aligning heart rate, IMU, audio, vision, dialogue, profile, and relationship graphs at the raw-signal layer to build a world model that genuinely understands the person, not a compressed version of general internet knowledge.

Roadmap

Three tracks running in parallel

Now — Always-on hardware → iOS → cloud main loop is production-stable; the full self-evolution loop is live; deep internal trial is underway.

Near term — Scale the self-evolution sandbox from small-batch precision tuning to industrial operation at the thousand-user public-beta scale, with quality gates moving into automation.

Mid term — In-house multimodal hardware v1 ships; the always-on device becomes a true external sensory system. End / edge privacy split goes live, making raw-data residency a user choice.

Long term — Three tracks converge: the Self-Evolving Agent delivers per-individual personalization at scale; the always-on hardware becomes a true external sensory system; the data flywheel matures into a self-trained Human-Centric World Model that understands people, not just the internet.

Join us

We're building a team that can lead long-horizon tracks end-to-end

Hardware, on-device multimodal, Realtime, Self-Evolving Agent, self-trained Human-Centric World Model — for every track, we're hiring people who can take full ownership end-to-end. Especially welcome: peers who already live with quantified-self, smart wearables, or personal AI.