Throughline

The problem

Direct support professionals (DSPs) support adults with intellectual and developmental disabilities in Medicaid home- and community-based services. DSP turnover was 37% in 2024, and 64% of the staff who left had been there under a year (NCI-IDD State of the Workforce 2024). Every departure takes with it what works for each person, and new staff fall back on the same group trip to the park or the mall. Federal HCBS rules require person-centered planning and real community integration, but care records only log what happened. They never suggest where to go next, or learn from how it went.

What it does

Throughline plans community outings that fit the person, learns from how each one went, and keeps the record when staff change.

Qloo proposes. An agent turns a person's loves (Selena, horses, tacos) into Qloo searches for real places nearby. Each search is filtered to places open that weekday, under the budget's price level, within travel distance, and with a step-free entrance where needed. Qloo's explainability shows which love drove each match.
Their reactions decide. After an outing, staff tap lit up, fine or not for them, plus optional tags like too loud or wanted to stay. Each log updates a per-person Beta posterior for every place feature (Qloo tags plus traits like quiet, outdoors, animals). Ranking blends Qloo affinity with what the logs taught, places marked not for them are excluded from future searches, and Thompson sampling picks one new place worth learning from.
The record survives turnover. The printable handoff card shows what lights them up, what to steer clear of, places to go back to, and how to support them. Every line comes from the profile or a logged outing, and progress-note drafts cite the log entries they come from.

What makes it Qloo-powered

Qloo supplies the taste prior that no care record has. We use /search to match each love to a Qloo entity, /v2/tags (semantic search) to match interests to tags and find the wheelchair-accessible-entrance tag, and /v2/insights for real places (signal interests and tags, location and radius, hours, price level, accessibility tags, excluded entities, quality bias, explainability) and for cross-domain at-home movie ideas on low-energy days. Throughline adds what Qloo can't know: how this one person actually reacted.

How we built it

Next.js 16, React 19 and TypeScript, with Tailwind and shadcn/ui. The planner is a tool-using agent on Claude (or on Gemini when only a Gemini key is set) with tools for place search, tag lookup, at-home search and a final plan submission, where every plan must reference a candidate a tool returned. Location, opening day, price ceiling and past exclusions are enforced in code, never left to the model. The agent sees the profile with the name removed, and Qloo receives only IDs, coordinates and filters. Staff watch each step live, including the exact Qloo request.

Try it

Open Maria, then Plan an outing, choose the Something new plan, then Log how it went with Not for them and Too loud. Watch the ranked list reorder and the "Learned from..." note appear, then open the Handoff card. The app labels which mode is running (Claude, Gemini, rules planner, or sample mode), and the seeded people and history are fictional.

Honest limits and what's next

Demo records live in browser local storage, and production needs per-agency storage with access control and audit logs. This has not been tested with a real agency yet. The next step is a two-week pilot with one agency and five people, measuring logging rate, new places tried, and time saved on notes. Qloo's place coverage is thinner in small towns, and libraries, community centers and events are not in its place taxonomy.