Gasp - Graph Agent Staffing Platform

Sub-Workflow 1
Sub-Workflow 3
Sub-Workflow 4
Sub-Workflow 2

Inspiration

Home-health agencies spend roughly 27% of their operating budget on non-clinical administrative labor, most of it manual phone coordination. The worst moment is the 6:00 AM call-out: when a caregiver phones in sick, a human scheduler loses about 45 minutes dialing backups one by one while billable shifts go unstaffed and patients wait. The same office also drowns in patient intake matching, end-of-day medication paperwork, and a language gap that forces them to pay outside medical interpreters by the minute. We wanted to know if a single voice AI could absorb that entire category of work, so a human never has to run the phone tree again.

What it does

The Healthcare Omni-Agent answers every inbound call for a home-health agency and handles the four jobs that eat scheduler time:

Shift-Backfill Cascade. A call-out triggers a Neo4j query for compliant, nearby backups, then fires an SMS and voice cascade. The first caregiver to say "YES" locks the shift live, turning a 45-minute scramble into about 5 minutes of oversight.
Intake and Graph Matching. New patient intakes are captured by voice and matched against the caregiver roster using graph-relational intelligence (certifications, languages, proximity), so the agency captures revenue faster and shrinks its waitlist.
Med-Log and Compliance Sync. An end-of-shift verbal questionnaire is turned into structured JSON and appended to an audit-ready log, removing manual compliance paperwork.
Interpreter Bridge and Escalation. A distress phrase hot-dials a human and opens two-way live interpretation, displacing external translation services that bill by the minute.

A Main Orchestrator Agent recognizes the caller by phone number, greets them by name, detects intent, and routes into the right workflow. If the caller switches topics mid-call, it closes one workflow and opens another without ever dropping the line.

How we built it

The system runs on a hub-and-spoke design: one Orchestrator (the hub) routes into four SubWorkflows (the spokes), with escalation as an always-on override.

The heart of it is a Dual-Agent core. An Active Agent owns the real-time audio stream, using ElevenLabs streaming TTS on a deliberately tiny context window so replies start fast. A Passive Agent listens to the WebSocket token stream in the background, writes to the database, and holds the global conversation state. On barge-in, Twilio halts playback, the Active Agent drops its in-flight generation and pulls the latest state snapshot from the Passive Agent, then resumes on the caller's new direction with prior context intact.

Twilio is the telephony backbone: Programmable Voice, SMS, Video, and Media Streams over WebSocket for inbound answering, the outbound cascade, and barge-in playback control.
ElevenLabs powers low-latency, natural voice for the Active Agent and the cascade calls.
Neo4j stores the graph-relational roster and serves the ranked-candidate and compliant-backup queries via Cypher.

We built the spine first (Twilio answers, ElevenLabs speaks), then the Dual-Agent core and barge-in, then the Orchestrator, then hung the SubWorkflows off it as thin vertical slices.

Challenges we ran into

Keeping voice latency believable. We had to stream STT and TTS and keep the Active Agent's context tiny, pushing every database write onto the Passive Agent.
Not losing state on barge-in. The interrupt path was the riskiest part of the build, so we snapshotted state in the Passive Agent and tested interrupt handling early.
Cascade race conditions. Multiple backups can reply at once, so we used a first-YES-wins lock and tested it under concurrent replies.
Graph matching in a time box. We seeded a small synthetic graph and kept the Cypher matching tractable rather than modeling the full domain.
Working with health data responsibly. We used only synthetic patient data and kept audit logs append-only.

What we learned

Splitting a live conversation into a fast responder and a background bookkeeper is what makes voice AI feel both quick and reliable. Once state lives outside the thing generating speech, interruptions and mid-call intent switches stop being scary. We also learned how much of an agency's cost is pure coordination, and how directly a single always-available agent maps onto real dollars saved per call.

What's next

Finish the Intake and Escalation workflows end to end.
Add the Twilio Video validation channel for field hazard checks.
Harden for real deployment: HIPAA-aware handling, richer roster data, and a live staff portal for the post-call summaries.