XingHan · built in-house

A model must understand elderly care to be usable on the floor.

XingHan is pre-trained on elderly-care text and fine-tuned on nursing records. It reads a medical order, and it reads the contracted shorthand found in a handover sheet.

Architecture
Mixture of experts (MoE).
Deployment
Device, edge and cloud.
Corpus
Continued pre-training on elderly-care text, plus supervised fine-tuning on care records.
Boundary
Private deployment; data does not leave the domain.
02

Capability

Medical orders, nursing notes and ward shorthand fall within its scope.

Beyond text, XingHan accepts voice and image input: a call recording or a single bedside frame becomes a structured record that can be searched and reviewed.

Text
Medical orders, care records, handover notes, ward abbreviations.
Voice
Day-to-day conversation and calls, transcribed into searchable text.
Image
Security and bedside frames, supporting safety-event assessment.
Knowledge
Coverage across four domains.
03

Device — edge — cloud

Judgement is made inside the building, not in the cloud.

80% of inference runs locally (target), and raw data stays on site. The cloud carries training and capability release only. When the external network fails, judgement inside the building continues.

Device
Bedside, wearables, access control and radar connect locally.
Edge
Inference on site; data does not leave.
Cloud
Continuous training; capability released by version.
Offline
Critical judgements do not depend on the external network.
04

Metrics and conditions

Every figure states its condition, including the unflattering ones.

All five figures below are lab-benchmark results or targets; the evaluation protocol is not yet published. Conditions are placed where readers can see them, because a figure without its condition carries no reference value.

Target 200ms latency (lab-benchmark condition; evaluation protocol pending).
Target 98.7% accuracy (lab-benchmark condition; evaluation protocol pending).
Target 80% of inference running locally.
Measured: F1 0.966 on supervised fine-tuning of care records (lab-benchmark condition).
Measured: F1 0.74 on cognitive screening (lab-benchmark condition).

Stage

The training pipeline is still in build, and its stage is stated as such.

CPT, SFT and RLHF are in build, with productisation planned for the second quarter of 2027. What can be demonstrated today is limited to the completed part.

Book a demo

Title

Thirty minutes to observe nine digital workers operating in a real care setting. An advisor will call to confirm the time and focus.