01FamNest
An AI wellness coach for exhausted parents. 2 parents in daily use.
Problem
Working parents of young kids are chronically depleted and have no time for "wellness" — and most apps only add to the load. The bar I set: ask 30 seconds of input and give back one small, doable thing that fits the day they're actually having.
Impact
Turned a single LLM call into a small multi-agent pipeline — a coach that drafts the plan, a safety reviewer that screens for crisis, a memory layer that tracks each parent's trends, and lightweight RAG that grounds advice in vetted guidance — so a 30-second daily check-in (mood, stress, sleep, time, goal) returns a warm, personalized micro-plan. Runs in production at zero API cost on a free model, with a deterministic safety floor for crisis cases that holds independent of the model.
Key Features
- Orchestrated multi-agent pipeline instead of one prompt: a coach agent drafts a history-aware plan around the parent's worst lever (sleep, stress, or mood), a safety-reviewer agent screens each draft (ok / revise / crisis) through a bounded revise loop, and a memory layer summarizes each parent's trends back into the next plan
- Deterministic crisis floor: a keyword guard escalates self-harm language to curated support resources, independent of the probabilistic model — real end-to-end testing caught two safety bugs a code review missed (the offline path had no crisis handling; the LLM classifier was non-deterministic on crisis)
- Lightweight lexical RAG grounds advice in a curated, sourced wellness corpus (chosen over vector embeddings — free and serverless-appropriate) with provenance on every plan; a provider-swappable LLM layer runs at zero cost on Groq's free tier, with a deterministic fallback plan when the model is down, all on Supabase RLS + rate limiting
