Inspiration
A lot can happen between one medical appointment and the next. A clinician may request labs, referrals, questionnaires, medication follow up, or home monitoring, while the resulting documentation arrives later from different sources.
The problem is that before the patient returns, staff still have to reconstruct what was requested, what is documented, and what still needs review.
While discussing the idea with a physician during the hackathon, he described how visit documentation is already becoming more automated, but said the workflow between visits could be more streamlined. That helped shape CareLoop around one idea: be the glue between appointments.
What it does
CareLoop is a pre-visit follow-up agent for primary-care staff.
It starts with the follow-up requests from the previous visit. As new synthetic records arrive between appointments, CareLoop interprets and stores them in the active chart.
Before the return visit, the agent:
- decides which record categories are relevant to each prior follow-up request
- retrieves the allowed evidence
- evaluates each request against that evidence
- identifies what is documented complete and what still needs review
- links findings back to the source records
- prepares a bounded draft staff handoff
CareLoop can also reassess the chart when new documentation arrives. For example, an ophthalmology item can move from Pending to Documented complete after a consultation note is received, while showing the new source that caused the review to change.
CareLoop does not diagnose, prescribe, determine treatment, or automatically contact patients.
How we built it
CareLoop is built with Python, Streamlit, OpenClaw, and structured local synthetic patient data.
The workflow separates AI reasoning from deterministic controls.
For incoming documents, an OpenClaw llm-task interprets the raw synthetic record. Python then validates the structured output, assigns a source ID, and stores the record in the active session chart.
For the pre-visit review, the agent first chooses which record categories are relevant to each prior-plan item. Python validates those choices and retrieves only the allowed evidence. A second model call evaluates the retrieved evidence.
Deterministic Python then verifies that every plan item is accounted for, statuses are valid, source IDs actually exist, and cited evidence came from the records retrieved for that item.
The interface presents the workflow as:
Previous Visit → Records Between Visits → Pre-Visit Review
It also preserves the previous review when new evidence arrives so CareLoop can visibly show what changed.
All patient information in the demo is synthetic.
Challenges we ran into
One of the biggest challenges was making the project genuinely agentic without giving the model unsafe control over a healthcare workflow.
We solved that by letting the model handle semantic tasks such as interpreting documents, selecting relevant evidence, and assessing documentation, while deterministic Python owns state, retrieval boundaries, provenance, allowed statuses, and staff-task generation.
Another challenge was distinguishing missing documentation from evidence that something did not happen. CareLoop deliberately keeps states such as No record found and Cannot verify separate and displays the reminder that not documented does not mean not done.
We also had to make the longitudinal behavior visible. Instead of showing only one generated review, CareLoop retains the prior assessment, marks it outdated when a new record arrives, and shows changes after the agent reassesses the chart.
Accomplishments that we're proud of
- Built a working end-to-end healthcare agent during the hackathon
- Added live interpretation of incoming synthetic records
- Implemented agent-directed evidence selection
- Added deterministic citation and provenance validation
- Made every finding inspectable against its underlying source records
- Built a before and after reassessment workflow when the chart changes
- Added a bounded draft staff handoff instead of autonomous clinical actions
- Kept the entire demo within a narrow and explainable safety boundary
One of our favorite examples is a medication documentation conflict: the current medication list and pharmacy record contain different directions. CareLoop does not decide which is correct. It surfaces both records and flags the discrepancy for clinician or pharmacist review.
What we learned
- Built a working end-to-end healthcare agent during the hackathon
- Added live interpretation of incoming synthetic records
- Implemented agent-directed evidence selection
- Added deterministic citation and provenance validation
- Made every finding inspectable against its underlying source records
- Built a before-and-after reassessment workflow when the chart changes
- Added a bounded draft staff handoff instead of autonomous clinical actions
- Kept the entire demo within a narrow and explainable safety boundary
One of our favorite examples is a medication documentation conflict: the current medication list and pharmacy record contain different directions. CareLoop does not decide which is correct. It surfaces both records and flags the discrepancy for clinician or pharmacist review.
What's next for CareLoop
Next, CareLoop could be tested with real pre-visit clinical workflows and expanded to support changing care plans, more document types, and persistent patient timelines. A future version could integrate with EHR systems so follow-up records automatically flow into CareLoop while preserving the same evidence-linked, human-in-the-loop review process.
Log in or sign up for Devpost to join the conversation.