Yasashii Bowel Care Agent

The problem

Family caregivers have so much to remember: what they observed, what needs recording, and what the next person needs to know. Bowel observations are part of that everyday workload. Repeated checking and remembering can add to an already demanding day.

Who it is for, and why it matters

Yasashii Bowel Care Agent is for family caregivers who share care with relatives or visiting professionals. The aim is to make records and handoffs easier while preserving the person's dignity and the caregiver's judgment.

AI helps. People decide.

What it does today

The prototype demonstrates an end-to-end guarded observation-to-handoff workflow using synthetic data. The demonstrated workflow starts with prepared observation inputs, not a live camera feed.

PASS — Record: the input passes the software's quality checks, and one observation is recorded. HOLD — Ask a person: the input is uncertain, so a request is saved for human review. It remains pending. STOP — Do not record an observation: the signal is unreliable, so the workflow saves a separate safety alert.

In the verified three-case demonstration, the handoff contains only the one PASS observation. The pending request and signal alert are kept separate. A read-only report makes these differences visible.

HOLD does not mean someone has approved the result. The caregiver approval screen is not implemented yet.

How Strands makes it work

The workflow uses the Strands Agents SDK:

Fixed rules check the input quality. A Strands agent selects a tool to record the observation, request human review, or stop. Executable guards check that the selected action is allowed before it can write a result.

The tools use the already checked input. They reject invented arguments, incorrect actions, and repeated calls. This keeps the model's choices within the workflow's rules. Inputs flagged as containing personal data stop locally before a model call.

The saved Mac run used Strands Agents SDK 1.55.1 + Ollama + qwen3:8b. No successful Amazon Bedrock run is claimed.

A basic bowel-monitoring prototype existed before the hackathon. The hackathon work adds the Strands workflow, PASS / HOLD / STOP policy, guarded tools, pending-review records, and handoff output. Reused work is disclosed in the repository. Codex helped with implementation, debugging, tests, evidence checks, documentation, and the bilingual demo.

What we verified

PASS / HOLD / STOP: all three synthetic cases completed in the saved Mac local-model run, with one matching tool action per case. One handoff observation: the PASS record is included; HOLD remains pending and STOP remains a separate alert. 13 source/input hashes matched: ten source files and three synthetic inputs matched the saved run's hashes. 109 automated tests passed: these use scripted models with network access blocked. They are separate from the saved Mac model run.

The demo video presents actual offline execution output and saved Mac AI-run results. The offline run does not call an AI model, and the saved Mac results are not live inference footage. The friendly guide is explanatory narration, not the agent's actual dialogue.

Large English text above smaller Japanese translations, clear labels, and symbols let viewers follow the main story with the sound off. The closing QR code and text link lead to the code, architecture, and evidence.

Limits

This is a synthetic-data prototype, not a finished care product.

Real camera integration and real-image detection performance are unverified. Real-care benefits and clinical safety are unverified. The caregiver approval screen is not implemented. Duplicate-write prevention covers one event execution; protection across process restarts remains future work.

The saved evidence covers one three-case model run. It does not establish broader real-world reliability. Hash matching checks file consistency; it is not an independent attestation of model execution.

This project supports observation, documentation, and communication. It does not diagnose illness or replace clinical judgment. An unobserved event does not prove that no bowel movement occurred.

What's next

Development is progressing toward private real-image evaluation with privacy safeguards. This is planned validation, not a completed result. We plan to compare outputs with human labels before making claims about real-image performance.

Further work includes camera integration, an explicit caregiver approval interface, duplicate prevention across restarts, and broader reliability testing.

Curious? Explore the code, setup, architecture, and reviewed evidence, including the saved Mac verification.

Technology that gently supports people.

Built With

• amazon-bedrock
• amazon-nova-lite
• jsonl
• ollama
• pytest
• python
• qwen3:8b
• strands-agents-sdk