FieldLine

Inspiration

I'm a researcher at KAUST on the Red Sea coast. When colleagues go out — reef stations by boat, desert sites by truck — the "safety system" is a filed PDF and a promise that someone will notice if they don't come back. At 18:05 on a Friday, nobody is watching that clock. A phone-call agent is the first technology that can actually work a safety plan: not send a notification into a muted group chat, but ring a human, ask the right questions, and climb a ladder of humans until one of them says "I'm taking over."

What it does

FieldLine turns a YAML trip plan into an active safety protocol. At each scheduled check-in, CALL-E calls the worker and brings back a structured result: safe / needs assistance / location / plan changes. Miss a check-in and FieldLine retries, declares you OVERDUE, then walks the escalation ladder call by call — briefing each contact with accumulated facts (last confirmed contact, vehicle position, what the previous contact reported) until someone explicitly assumes coordination. Every incident ends in a written brief: timeline, transcripts, evidence, recommended actions.

The wow: a pre-agreed duress phrase. Say it mid-call and the agent does not react — it ends the call normally, then silently escalates straight to the safety officer with instructions not to call your phone back. FieldLine also never auto-dials emergency services; the last rung is always a human who does.

How we built it

Every conversation is one CALL-E call task: POST /v1/calls via the official calle-ai Python SDK, with a natural-language task and a JSON result_schema. CALL-E plans the call, dials, talks, and returns structured_result, task_completed, completion_confidence, and evidence — FieldLine's decisions consume only those fields. The silent-duress protocol is pure prompt + schema: the task instructs the agent to never react to the phrase, and duress_phrase_detected: boolean rides home in the structured result.

Above the SDK sits a deliberately deterministic core: protocol.py is a unit-tested state machine (32 tests) — no LLM in the safety loop, so escalation decisions are reproducible. The dispatcher wraps every SDK call with timeout + one retry made double-dial-proof by idempotency keys, and fails soft: if the telephony API itself is unreachable, the synthetic failure is classified as "could not reach" and the cascade keeps climbing instead of crashing — an outage degrades into exactly the behavior a safety system should have. A demo transport replays scripted calls shaped byte-for-byte like the OpenAPI call_task schema, so the entire product runs offline with zero keys and going live is a transport swap.

Challenges

Designing duress handling was the hard part: the one thing the agent must not do is acknowledge the phrase, so detection had to live entirely in the extraction schema rather than in the conversation. The second challenge was failure semantics — for a booking bot a dead API is an error message; for a safety net it must be information ("the worker could not be reached"), which forced the fail-soft synthetic-result design. And keeping the demo honest took discipline: the simulated transport is validated by shape-conformance tests against the real schema, and every fake is marked // DEMO in source.

Accomplishments

A complete safety protocol — schedule, retries, grace windows, silent duress, stand-down rules, ladder exhaustion — as a solo build inside the window, with 32 passing tests, a consent gate before any live call, cancellation honored between calls, masked phone numbers everywhere, and an incident brief a security office could actually act on.

What we learned

An agent that acts in the physical world needs its judgment split in two: let the platform be smart on the call, and keep the policy around it boring, deterministic, and testable. Also: for safety software, "what happens when our own vendor is down" is a feature, not an edge case.

What's next

Publish the check-in as a CALL-E Goal so organizations can reuse it; webhook-driven async monitoring of many workers at once (a fleet board for field-safety officers); an SMS rung between phone rungs; and pilot it with an actual university field-safety office — the trip-plan PDF they already require contains every field FieldLine needs.

Built with

python · uv · calle-ai (CALL-E Python SDK) · CALL-E /v1/calls + result_schema structured extraction · rich · pyyaml · pytest

Disclosures

Built during the submission period with AI coding assistance (Claude); design, review, and verification by the author. Demo-mode transcripts are hand-written simulation data, marked // DEMO in source; live mode uses the real CALL-E SDK unchanged.

Built With

• call-e
• calle-ai
• pytest
• python
• pyyaml
• rich
• uv