nb there are more videos on the youtube channel.
Inspiration
Watching someone you love slowly deteriorate is one of the hardest things you can experience. Recent advancements in dementia care have been able to slow this deterioration and provide massive improvements to quality of life. But in practice, the personalised approach they require is impractical to deliver on a large scale, and impossible with our constrained healthcare resources.
With AI, we can assist carers in delivering truly personalised care, reconnecting patients with their precious memories and provide them with the anchor they need.
What it does
Anchor is a clinically grounded AI companion for carers. Whilst the patient is able to recollect, the AI conducts warm, open-ended conversations to build a rich personal memory profile: key people, photos, meaningful places, favourite music, life stories, and personality traits. With support for photos and real-time voice and video, patients can interact in a way that feels natural to them. The AI collects structured data on the patient, which is presented for the carer to review. In the early stages, this recollection is helpful in grounding the patient.
However, as the disease progresses, being asked to recollect becomes painful and causes distress and agitation. Anchor uses the data collected to provide personalised comfort sessions, gently narrating the patient's memories back to them. By offering memories, rather than demanding recall, the patient is able to recall at their own pace and ability. The AI automatically uses sounds and photos to activate non-episodic memory, which is proven to be effective in restoring functioning for patients with dementia. Music is suggested by the AI, but links are entered and supervised by the clinician to avoid incorrect links and adversarial attacks.
Crucially, the entire process is supervised. The AI never quizzes, corrects or impersonates, alerting carers automatically when a human touch is needed. All data is stored locally on device, and where data is unable to be sent to cloud model providers for legislative or privacy reasons, these could be substituted for local models.
One might ask, does this replace carer interaction. No. This tool is designed as an aide, with the process providing data to both the carer and family, whilst the treatment itself should enable patients to maintain their memories for longer, improving the quality of interactions with family members.
How we built it
Python, FastAPI, SQLite db. Claude handles messages and photos, whilst gemini-live handles the multimodal video and voice modes. The frontend is in vanilla JS, HTML/CSS. Simple with large text, high contrast and a calm visual design appropriate for the clinical context.
Challenges we ran into
Dementia requires a restrained warmth that avoids confrontations or correcting delusions, whilst avoiding sycophancy. The text mode attempted to use *** actions *** descriptions typical of internet text and engaged in sycophantic behaviour, which needed to be tuned out. Negative instructions for what the AI should never do proved more effective than prescribing what it should.
The entire review flow is built with carers in mind. All data is supervised by humans, with a diff and confirm step to highlight new items they might miss. The app can be used to aid carer interaction by providing them with the data and transcripts from the AI's interactions.
Accomplishments that we're proud of
A system prompt grounded in real clinical frameworks: Validation Therapy (Naomi Feil), person-centred care (Tom Kitwood), and research on musical memory preservation in Alzheimer's (Jacobsen et al., 2015).
Stage-adaptive AI behaviour — the AI genuinely changes its approach across early, middle, and late stages, shortening responses and shifting toward music and simple affirmations as the disease progresses
Emotional tracking - The AI can aid carers in tracking progression by assessing the patient's emotional state through their interactions.
Log in or sign up for Devpost to join the conversation.