CareLM · Senior-safe base model
A customized GPT-OSS base, senior-safe by training.
CareLM is an open-weights GPT-OSS base model, further trained for safe, adequate senior voice interaction. It is tuned for the conversation dynamics of aging — under medical committee supervision, with locked versioned weights for reproducibility.
GPT-OSS baseGeriatric-tunedMedical committeeLocked weights
Training inputs
Trained on real geriatric conversation, under supervision.
Every input is de-identified and anonymized, and the corpus is assembled under medical committee supervision.
De-identified Amigo data
Anonymized and de-identified conversation data from the Amigo companion platform — real senior interaction, stripped of identity at the source.
DementiaBank
Established research corpora of speech from older adults, including populations with cognitive impairment, grounding the model in clinically relevant language.
Proprietary geriatric corpora
In-house datasets curated for geriatric conversation dynamics, assembled and reviewed under medical committee supervision.
Why a dedicated model
Generic LLMs mishandle the moments that matter most.
Off-the-shelf assistants are tuned for productivity and speed. Geriatric conversation asks for the opposite — patience, consistency and care through repetition, confusion, distress and slower pacing.
Repetition
Generic assistants get impatient or reword; CareLM meets repeated questions with steady, consistent warmth.
Confusion
Rather than correcting or overloading, CareLM slows down, reorients gently and keeps the person grounded.
Distress
Signs of anxiety, panic or upset are handled with de-escalation and escalation awareness, not scripted deflection.
Slow pacing
Turn-taking, pauses and speaking rate are tuned to the person, not to an efficiency-optimised chatbot cadence.
Safety behaviors
Safety is trained into the weights, not bolted on.
The behaviours that keep a senior interaction safe are part of the model itself, and pinned to a specific version.
Identifies as AI
CareLM states that it is an AI companion. It never impersonates a human, a clinician or a family member.
Escalation-aware
Trained to recognise when a conversation should surface an alert or hand off to a caregiver, feeding the VGDS safety path.
Distress handling
De-escalation behaviours for panic, dissociation and acute upset are built into the model, not bolted on downstream.
Locked, versioned weights
Behaviour is pinned to specific weights so a deployment is reproducible — the same version behaves the same way every time.
Deployment
Runs where you need it — and licensing is available.
CareLM ships inside Amigo, inside bespoke platforms we build, and on-premise in your own infrastructure. The model can also be licensed directly.
ProductLearn more ->
Inside Amigo
CareLM is the conversational core of the Amigo companion — every senior call runs on it.
ServiceLearn more ->
In bespoke platforms
The model powers proprietary voice agent platforms we build around a senior care company's operations.
ServiceLearn more ->
On-premise
Runs inside your infrastructure for data sovereignty, with white-label Amigo or a bespoke deployment.
Continuous improvement loop
A supervised loop from conversation to locked release.
Data feeds supervision, supervision gates training, training is evaluated, and only then does a version release — keeping every deployment safe and reproducible.
01
Data
De-identified Amigo conversations and proprietary corpora feed the pipeline, anonymized at the source.
02
Supervision
The medical committee reviews inputs, behaviours and edge cases before anything reaches training.
03
Training
The GPT-OSS base is further trained for geriatric conversation dynamics and safe senior interaction.
04
Evaluation
Candidate weights are evaluated for safety behaviours, pacing and clinical adequacy before release.
05
Release
A version is locked and pinned; deployments upgrade deliberately, keeping behaviour reproducible.
CareLM is a base model for safe senior interaction. It is not a medical device and does not provide diagnosis or medical advice.
Talk about CareLM licensing.
License the model, deploy it on-premise, or build a bespoke platform on top of it. Tell us the deployment context and we will map the route.