Eastwood takes purchased or inbound leads, decides who may lawfully be called and who is worth calling, then has Imogen, a realtime AI voice assistant, run the discovery call. She screens health, presents deterministic quotes, and books the prospect with a licensed human agent who closes. Every step is gated by code, not by asking the model nicely.
Each lead passes through the same pipeline. The green steps are hard gates: if one fails, the lead stops there, and nothing downstream can override it.
CSV or PDF lead exports with fuzzy column mapping, an HTTP ingest API, signed Facebook Lead Ads webhooks, and web forms.
Phones to E.164, product lines, age from DOB, and consent records. Unmapped columns are kept for audit.
TCPA 8am–9pm local window, National DNC clearance (fail-closed), prior opt-outs, express written consent.
Medical knockouts, severity scoring, wealth and fit signals, then a route: auto-dial, request consent, nurture, or suppress.
A concurrency cap, a 4/24/32/56/80-hour retry ladder, max attempts, and idempotent handling of Twilio status callbacks.
Discloses she is an AI, confirms identity, runs needs discovery and health screening, and captures only what she clearly heard.
Unlocks only after discovery is complete. The engine picks three options; the model only explains them.
A real Calendly booking, a full prospect profile persisted, and an SMS brief to the licensed agent. Outcome-aware follow-ups run after.
Speech-to-speech models predict the likeliest next line of a sales call, which is "Perfect, thanks" followed by the next question, even when the prospect only coughed. Eastwood takes the decision to speak away from voice detection and rules on what was actually said first.
Every answer moves UNKNOWN → HEARD → CONFIRMED. Only confirmed values reach the quote engine or the handoff brief. "Unknown" is an acceptable outcome; a wrong value is not.
To store a value, the model must quote what it heard. That quote is matched against finalized transcripts of the prospect's own speech and scored with per-token logprob confidence.
Purpose → beneficiary → burial or cremation → coverage → rating sex → tobacco → health. Code owns which question is next; the persona owns the wording.
Barge-in cancels speech and flushes queued audio. A watchdog nudges a stalled model, then ends the call cleanly. After three unheard turns she names the line problem instead of looping.
Grouped by subsystem. Every item below exists in the code today; the core paths are covered by the 453 automated tests.
POST /api/leads/ingest HTTP APIauto_dial, request_consent_sms, nurture, manual_review, or suppressCALL_NOW / WORKABLE / DEAD buckets, a versioned rulesetupdate_lead_field, add_followup_questionsget_quote, select_quote_optionget_booking_slots, book_appointmentcapture_consent, capture_objection, capture_dnctransfer_live, transfer_to_human, end_callend_call remains.quote_type, rate_source, carrier_rate, bindableagent_policy_decisions_totalhandoff_profilesThe right column matters as much as the left. Most of these refusals are enforced in code, at the point where the action would happen.
clear or exempt on the DNC registry, or without express written consent.Everything runs from a single docker compose up. Only signed Twilio webhooks and the media WebSocket are exposed at the edge; health checks, metrics, and session admin stay internal.
Ingest, normalization, triage, scoring, routing, DNC/TCPA gates, placement, quoting, booking, dialer loop, follow-up worker, notifications, operator APIs.
Imogen: the Twilio ⇄ OpenAI Realtime bridge, turn gate, discovery flow, slot state, tool gating, policy engine, call tracing.
The application code is complete for a first controlled campaign. What remains is configuration, credentials, and data. That work is tracked in the go-live checklist.