ParentCare
Kids assign tasks. Parents complete them. AI verifies the proof.
Solo build: full-stack, architecture and AI
Adult children move to cities and abroad while their parents age at home. A checkbox 'I took my medicine' is too easy to fake, so nobody trusts it and nobody actually follows through.
Turn care into verifiable proof. Kids create recurring health tasks; parents snap a photo (a plate of food, a medicine strip); an AI agent named Saathi verifies it with vision and streams a plain-language verdict (passed, flagged, or failed) back to the kid's dashboard in real time, with streaks, badges, and a weekly report.
Under the hood
Async AI verification pipeline
Photo submissions persist and return 201 instantly, then fire-and-forget to a worker. A 10 to 30s Gemini Vision loop runs off the request path, modeled as an explicit task state machine (pending, submitted, then passed, flagged or failed) so serverless timeouts never hold an LLM open and elderly users never watch a spinner.
Agentic loop with 16 tools
The verification agent runs a real Gemini tool-use loop: read tools (history, nutrition trend, health profile) and action tools (update result, flag concern, alert kid, generate meal plan or exercise routine, read medication label). Four distinct agent loops, every run audit-logged, with a safety net that force-flags if the model never returns a verdict.
Queue with a zero-infra fallback
Runs on BullMQ and Upstash Redis (retries, backoff, persistence) in production, and inline in-process when Redis isn't configured, with an identical HTTP contract both ways, so a fresh clone runs with no infrastructure.
Security enforced in the database
Row-Level Security on 20 of 25 Postgres tables with per-role policies, Storage RLS scoped to a photos/<parentId>/ path prefix, and a SECURITY DEFINER trigger that mirrors every auth signup into the app schema atomically.