Inspiration
CareLink AI was inspired by a gap we observed in the current healthcare and caregiving ecosystem. In Singapore, many elderly residents, especially low-income seniors or those living alone, depend on free or subsidised care services such as home visits, physiotherapy, medical transport, meal delivery, and befriending programmes.
The problem is not that these services do not exist. The real issue is that after a volunteer or healthcare worker identifies a resident's need, there is often limited visibility on whether that need is actually followed up. For example, a volunteer may visit an elderly resident, notice joint pain or mobility issues, and refer the case for physiotherapy. However, because the handoff is often informal and dependent on volunteer follow-up, referrals can fall through the cracks.
Existing platforms such as Homage and Elderwise address parts of the caregiving problem, but they are mainly built for paid care or professional caregiver contexts. CareLink AI focuses instead on the free, volunteer-driven layer of community care by providing a coordination and communication platform that strengthens the links between non-profit organisations (NPOs) as service providers, caregivers, and care recipients. Built on the core belief that community care should be measured by real follow-through, not just referrals made, CareLink AI helps NPOs track whether identified needs are followed up, resolved, and meaningfully met.
What it does
CareLink AI is a healthcare ecosystem platform that strengthens the link between caregivers, service providers, and care recipients.
The platform allows volunteers or healthcare workers to log resident needs after home visits. AI then helps convert informal notes into structured care categories such as mobility support, medical follow-up, social isolation, or financial assistance. This makes it easier for nonprofits and service agencies to understand the resident's situation and take action quickly.
CareLink AI also supports service matching by recommending suitable providers based on eligibility, location, cost, and availability. More importantly, it includes service gap detection. If a resident is supposed to receive a scheduled service, such as weekly physiotherapy, but no visit is logged for two consecutive weeks, the system automatically flags the case and prompts follow-up, reassignment, or escalation.
The platform also includes role-based views. Volunteers and agencies can view fuller case notes and internal flags, while residents and caregivers see a simpler version showing what is scheduled, completed, or pending. This keeps the system transparent while protecting resident dignity and sensitive case information.
Please note that the AI features in the live app are currently limited to demo mode due to .gitignore file exclusions. The live app does not currently reflect the full capabilities of CareLink AI. For the most accurate representation, please refer to our video demo.
How we built it
We designed CareLink AI around three main user groups: residents, volunteers or healthcare workers, and service providers.
Residents can register or be referred through a community organisation, with Singpass verification used to support identity and eligibility checks. Volunteers join through invite codes issued by their sponsoring nonprofit or agency, so the platform builds on existing volunteer vetting instead of creating a completely new trust system. Service providers are platform-onboarded through a pre-approved list, since provider verification is closer to accreditation than normal self-registration.
For the prototype, we planned a Streamlit-based demo with three main role views. The backend uses structured resident, volunteer, provider, and visit data. The most important feature is the service gap detection logic, which tracks whether scheduled visits are actually logged. We also use an LLM to process informal volunteer notes and convert them into structured JSON categories for easier case management.
We kept the technical scope realistic. Needs assessment is LLM-driven, service matching is AI-ranked filtering over structured data, and gap detection is state-tracking logic. We deliberately placed predictive risk detection as future work because that would require more longitudinal data and model validation than a hackathon prototype can honestly support.
Challenges we ran into
One major challenge was avoiding overclaiming. It would have been easy to describe the platform as a fully predictive AI healthcare system, but that would not be technically honest. We had to separate what AI can realistically do now from what should remain as a future roadmap feature.
Another challenge was designing for multiple stakeholders with different needs. Volunteers, agencies, service providers, residents, and caregivers should not all see the same information. A shared backend is useful, but a single shared dashboard could expose sensitive notes or create confusion. This pushed us to design role-based views instead.
We also had to think carefully about trust and misuse prevention. Since the platform deals with vulnerable residents, it cannot be treated like an open marketplace. Volunteers need organisational accountability, residents need accessible onboarding, and service providers need proper approval before being recommended.
Finally, our weakest area is evidence of real demand. While the problem is plausible and grounded in Singapore's care context, we still need stronger validation through user interviews, nonprofit feedback, or local data on referral follow-through and missed service delivery.
Accomplishments that we're proud of
We are proud that CareLink AI does not just add AI for the sake of sounding innovative. The AI is used where it solves a real workflow problem: turning messy volunteer notes into structured needs and helping agencies prioritise follow-up.
We are also proud of the core reframe behind the project. Instead of asking whether a referral was made, CareLink AI asks whether the resident actually received help. This makes the product more outcome-focused and better aligned with the real purpose of community care.
Another accomplishment is the service gap detection feature. It is simple but powerful because it creates accountability after a referral is made. A missed visit no longer disappears quietly. It becomes a trackable case that can be followed up, reassigned, or escalated.
We also developed a clearer onboarding model for different users. Residents, volunteers, and providers each have different trust problems, so they should not all be onboarded in the same way.
What we learned
We learned that the strongest AI products are not always the ones with the most advanced models. Sometimes, the more valuable solution is a practical system that combines AI with good workflow design, accountability, and clear user roles.
We also learned that product scoping matters. A hackathon solution must be ambitious enough to be meaningful, but realistic enough to be built and defended. Moving predictive risk detection to the roadmap made the product stronger because it showed technical discipline rather than exaggeration.
Most importantly, we learned that community care should be measured by outcomes, not administrative activity. A referral is only useful if it leads to actual support. CareLink AI is built around that principle.
What's next for CareLink AI
The next step for CareLink AI is to stress test the platform with a selected group of relevant stakeholders, including volunteers, non-profit organisations, caregivers, and care recipients. This pilot would help us validate whether the platform fits real community care workflows, whether the role-based portals are easy to use, and whether our AI-supported features genuinely reduce manual coordination work. Feedback from this testing phase would guide improvements to the platform’s usability, accuracy, and follow-up process.
We would also strengthen the prototype by building a more complete end-to-end care journey supported by clearly scoped AI features.
For resident assessment, an LLM would convert informal volunteer notes into structured JSON fields such as need category, urgency level, suggested service type, and confidence score. Low-confidence outputs could be flagged for human review instead of being automatically accepted.
For service matching, the platform would use AI-assisted ranking to compare resident needs against a structured provider database, considering factors such as eligibility, location, cost tier, language, availability, and service type.
For volunteer coordination, a weighted matching system could recommend suitable volunteers based on proximity, language, skills, availability, and case urgency. Missed-visit detection would remain a state-tracking feature rather than a prediction model, comparing scheduled visits against logged visits and triggering escalation rules when care delivery lapses.
In the future, CareLink AI could expand beyond individual case management into community-level insights. By analysing repeated needs across multiple residents, the platform could help nonprofits identify broader patterns across a block, school, or neighbourhood. For example, recurring mobility issues, food insecurity, social isolation, or missed medical appointments could signal a wider community-level problem. This would allow organisations to move from reacting to individual cases toward planning more targeted community interventions.
CareLink AI could also support more specialised stakeholders within the community care ecosystem. One possible direction is a skill-based volunteering module that connects corporate volunteers, NPOs, and care recipients more effectively. Instead of matching volunteers only by availability, the platform could use AI-assisted matching to consider skills, language, location, lived experience, and the type of support required. This would make volunteering more purposeful while helping NPOs access the right kind of support for different care needs.
Longer term, CareLink AI could introduce predictive risk detection once enough reliable check-in data, service history, and outcome feedback has been collected. Instead of making clinical decisions, this model would act as an early warning layer for human reviewers. It could analyse signals such as repeated missed visits, unresolved needs, negative feedback, frequent reassessments, social isolation indicators, and delays between referral and service completion. Over time, the system could learn which combinations of signals are associated with cases requiring closer follow-up. However, this feature should only be deployed with consent, human oversight, audit logs, and validation against real community care data, especially because the platform serves vulnerable residents.
Log in or sign up for Devpost to join the conversation.