It phones every doctor's office on a health plan's list, fixes the list, and keeps the clip behind every change.
Live: https://rollcall-sage.vercel.app · slides · how it is built
Built on the AssemblyAI Voice Agent API for the AssemblyAI Voice Agent Hackathon (lablab.ai, September 2026).
Every office in this project is simulated. No real clinic was called, and every phone number is in the 555-01xx range reserved for fiction.
US Senate Finance Committee staff called 120 mental-health listings from 12 Medicare Advantage directories. A third were wrong or unreachable and they could book an appointment 18% of the time. The No Surprises Act already requires plans to verify every listing every 90 days; the REAL Health Providers Act makes that proactive from plan year 2028. Today that is a call centre asking the same five questions.
- Sweep. One click. Six lines dial 40 offices. Each call checks five facts in three questions: still here; address and booking phone; new patients and in network.
- Write, with evidence. Every clear answer becomes one tool call. Confirmed cells get a tick; corrected cells show the old value struck through beside the new one. Click any cell to hear the receptionist say it.
- Refuse to guess. "I think he's still taking patients?" is not written down. It goes to a person, with the clip.
- Your turn. Answer a call in the browser and play the front desk. What you say becomes row 41.
- Report. A printable attestation sheet and a corrections CSV.
| AssemblyAI feature | What it carries here |
|---|---|
| Voice Agent API, one WebSocket | Both ends of every recorded call (Rollcall and the simulated front desk), and the visitor's live call from the browser |
input.keyterms + transcription_prompt |
Built per call from the row being verified: the provider's surname, the practice, the street, the plan (lib/agent/session.ts) |
| JSON-schema function tools | Seven tools; each clear answer is exactly one call (lib/agent/tools.ts) |
tool.result with is_error |
The write gate's refusals go back to the model as instructions ("'2310' was not heard. Ask them to repeat the street number.") |
| Turn detection and barge-in, left on adaptive defaults | Front desks interrupt, say "hold please", and read numbers slowly |
Temporary tokens (GET /v1/token) |
The browser never sees the API key; each token caps the session at 180s (app/api/token) |
reply.create |
If the visitor picks up and says nothing, the agent says hello first |
No greeting |
The office answers first, as on a real call |
execution_mode per call type |
hold for recorded calls (faster), interactive with a spoken "Got it." for live ones (no dead air) |
Voice-gated input (lib/agent/voice-gate.ts) |
Audio is sent only around speech; a continuous stream measurably delays turn detection (docs/adr/0005) |
The write gate (lib/agent/gate.ts) is code, not prompt. A value reaches the directory only if the office's own words support it: numbers in a new address or phone must have been heard (in any spoken form), a hedged answer is refused, the second refusal sends the field to a human, and the call's outcome is derived from what was written (lib/agent/outcome.ts), never taken from the model.
flowchart LR
R[Directory row] --> K[key terms + prompt from the row]
K --> A[Voice Agent API call]
A -->|tool.call| G{write gate}
G -->|office said it| C[cell + audio span]
G -->|hedged / not heard| E[tool error → agent re-asks]
E --> A
G -->|second refusal| H[human review]
C --> O[outcome derived in code]
- Buyer: the provider-data team at a health plan. The work is mandated every 90 days and is done today by phone.
- Cost per listing is computed from the calls themselves and shown on the console (about $0.05 at the $0.075/min Voice Agent rate for a 43-second call).
- Prior art: enterprise vendors sell this to large plans, which validates the buyer. Rollcall's wedge is the evidence trail and the refusal to guess.
- Next: the same engine fills any column that only a phone call can answer.
Most voice agents replace a receptionist. This one calls them, many at once, and treats every answer as a claim that needs a receipt.
The console is a register, a wall map and a six-line desk phone (DESIGN.md, UI-SPEC.md). /_kit shows every component in every state.
| Criterion | Where to look |
|---|---|
| Application of Technology | lib/agent/ (tools, session builder, write gate, live driver), scripts/lib/bridge.ts, ADRs 0001, 0005, 0006 |
| Presentation | the landing page at / (live console in the hero), the console at /console, every component and state at /_kit, DESIGN.md, UI-SPEC.md |
| Business Value | /about (the rule, the benchmark, cost per listing), /report (the artifact a plan would file) |
| Originality | click a struck-through cell; open Needs a human; open any transcript with a red "Refused by the gate" row |
| Does it work? | npm run verify (no credentials), npm run eval (40/40 outcomes, 150/150 written fields, 0 false writes), tests/replay.test.ts replays every recorded call through the gate |
| Real | Simulated |
|---|---|
| The agent, its tools, the write gate, outcome derivation, the live browser call | Every office, provider, address and phone number |
| All 40 calls on the console: recorded through the Voice Agent API, with their audio | The health plan and the county |
npm run verify: 121 tests + 12 checks over the recorded data, no credentials |
The front desks are language models playing scripted offices; real ones have phone menus and hold queues |
Scored against the offices' hidden truth sheets (npm run eval): 40 of 40 outcomes, 150 of 150 written fields, 0 false writes; the gate refused 5 tool calls. Three listings were re-recorded after the bug they exposed was fixed in code; the other 37 are first takes. Each of those bugs is now a test built from the real call.
npm install
npm run dev # http://localhost:3210 (landing) and /console, no key needed
npm run verify # typecheck + tests + judge-check, prints PASS/FAILWith a key (ASSEMBLYAI_API_KEY=... in .env.local): npm run check:keys, then npm run record, npm run import:recordings, npm run eval.
- English only. One network, one call script.
- The simulated front desks are language models too; real offices have phone menus and hold queues this does not navigate yet.
- The live-call budget limiter is in memory per server instance. The app does not rely on server-side session caps: the UI ends every call at 150 seconds, budgets are low, and the account runs on prepaid credits.
- A visitor's call audio lives in the browser tab and is gone after a refresh; the row and transcript persist.
Built with Claude Code. Design and build decisions are recorded in DESIGN.md, UI-SPEC.md and docs/adr/. MIT licensed.