Inspiration
In clinical research, a missed visit confirmation is not a soft calendar problem. It is a wasted protocol slot, a delayed endpoint, and another hour a coordinator spends in voicemail limbo.
Coordinators already juggle SMS blasts and one-at-a-time appointment tools. What they still lack is a batch phone confirmation rate they can act on before clinic day — who said yes, who needs a new slot, who never picked up — with research-visit consent language on the line.
VisitLock started from that ops gap: treat research-visit confirmation as measurable phone work, not inbox archaeology.
What it does
VisitLock batch-confirms study visit slots from a CSV. Each participant gets a phone confirmation (live CALL-E or offline fixture) with explicit research-visit consent language. The agent must return structured RSVP fields: visit_status (yes | no | reschedule | no_answer | unknown), preferred_slot, and notes.
An idempotent ledger keyed by participant_id::visit_datetime prevents double-dialing when a batch is re-run. The coordinator HUD shows confirmed/called, confirmation_rate %, and reschedule_count. MATLAB quotes confirmation_rate and a no_show_risk score onto the same board so risk sits next to the confirmation KPI — not buried in a notebook.
Judges (or anyone) can open the public fixture HUD with non-zero numbers — no API key and no localhost:
https://moscraciunxxx.github.io/visitlock-calle/
Fixture snapshot: 3 / 5 confirmed, 60.0% confirmation rate, 1 reschedule, MATLAB no_show_risk 10.0.
Why this is not a generic appointment bot: batch research-visit CSV, research-consent wording, idempotent dial ledger, confirmation_rate as the ops metric, and MATLAB no-show risk on the same board.
How we built it
- Python CLI (
python -m visitlock) for demo, CSV run, serve, and docs export. - CALL-E via
calle-aiCalleClient.calls.create_and_waitwhenCALLE_API_KEYis present; otherwise a deterministic fixture that preserves the same schemas. - JSON Schema on call results so RSVP status is machine-checkable, not free text.
- Static
docs/index.html+docs/board.jsonpublished on GitHub Pages as the stranger-openable door. - MATLAB script stamps metrics into
artifacts/matlab/and the board JSON. - Public Pages HUD is the board source of truth.
- Contribution pack for the awesome list: skill + app pointer (PR #333).
Challenges we ran into
Keeping the public story honest was harder than wiring the happy path. The live door had to work cold on Pages with fixture numbers baked in, while live dials stayed behind a gitignored .env. We also had to draw a clear line versus generic appointment-confirm tools: batch study visits, ledger, confirmation_rate HUD, and research-visit consent language. Matching MATLAB-quoted risk to what the HUD shows — and keeping the public board as the only source of truth — took more editing than expected.
Accomplishments that we're proud of
- A stranger can open the Pages HUD and read real fixture metrics in under thirty seconds.
- Fixture and live paths share the same RSVP schema and board shape.
- MATLAB risk is visible next to confirmation rate, not buried in a notebook.
- Awesome-list skill/app PR opened with an upstream pointer to this repo.
- Sample phones stay in the fictional 555 range; no secrets in the tracked tree.
What we learned
Phone agents for research ops need structured outputs more than clever dialogue. Coordinators care about confirmation_rate and who still needs a human callback. Shipping a static fixture door early forced the product story to stay readable without a key, and it made live CALL-E a deliberate second step instead of a demo blocker.
What's next
- Upload and link the demo video on Devpost (local cut lives outside git).
- Tighten live-call safety: clearer destination checks and base-URL defaults before
--live. - Optional coordinator actions from the HUD (retry no_answer, accept reschedule slot).
- Broader fixture scenarios (multi-site, multi-day) once the core board stays stable.
Log in or sign up for Devpost to join the conversation.