Senior Telephony Engineer (Telnyx Call Control) to build production-grade voicemail detection for our live transfer

About us
Myra is a sales operations platform for agencies and their clients: lead capture, automations, calendars, messaging, and a calling system with live transfers. When a new lead comes in, Myra phones them within seconds, plays a short greeting, and connects them to a sales agent. Live transfer works well and is in daily use by paying customers.

The problem we need solved
Voi cemail s. When the lead's phone goes to voi cemail , our agents get connected to a recording instead of a person. We have had answering machine detection (AMD) switched off because every version we tried was not reliable enough: Telnyx's built-in AMD contradicted itself on roughly a quarter of calls, and once we made the greeting play instantly on answer, the AMD verdict arrived too late to act on. We have spent two months on this without getting it to production quality. We want someone who has solved this before to own it end to end.

What "done" looks like

Voi cemail s almost never reach an agent (target: under 3% of transferred calls).
Real people are almost never hung up on by mistake (target: under 1%). This matters more than the first target: when unsure, connect the human.
The decision is made within about three seconds of answer, without the lead hearing dead air.
It is measurable: every call logs the signals used and the decision, so accuracy can be reported weekly against what actually happened on the call.
It is safe to roll out: shadow mode first (decide and log, but take no action), then enabled per customer behind a flag, then default on.
Documented and handed over so our team can maintain it.

Our stack

Telnyx Call Control (voice), Telnyx messaging
TypeScript throughout: TanStack Start on Cloudflare Workers, Supabase (Postgres, edge functions, pg_cron)
Built and deployed via Lovable with the repo on GitHub; you will work through pull requests reviewed before publish
Call recordings and transcripts available for evaluation

What you will do

Review the current live transfer flow and the previous AMD attempts (we have the data on what went wrong).
Design the detection approach. We have a preferred direction (a short prompt to the lead, speech detection on the reply, Telnyx beep detection and real-time transcription as signals, AMD as a hint only, a scoring decision with a hard time limit that defaults to human) but we want your judgement.
Build it behind a feature flag with full instrumentation.
Run it in shadow mode on real traffic, measure against recordings, tune until the targets are met.
Roll it out gradually and hand over with documentation.

You must have

Deep, hands-on experience with Telnyx Call Control (or Twilio Programmable Voice with a clear plan for Telnyx): call legs, bridging, media, webhooks, timing.
Prior production experience with answering machine or voi cemail detection, and honest numbers on how accurate it was.
Strong TypeScript/Node, Postgres, and webhook-driven systems.
Production discipline: feature flags, logging, shadow rollouts, and the instinct to never break a working call path.
Comfortable working in an existing codebase you did not write.

Nice to have

Real-time speech-to-text (Deepgram, AssemblyAI, Google STT) over streaming audio.
Cloudflare Workers and Supabase.
Experience with dialler or contact-centre products.

Engagement

Contract, exclusively on this feature; no other tickets.
Working hours: 7 AM to 12 PM Pacific, Monday to Friday, with daily check-ins.
Expected duration: 4 to 6 weeks including the shadow-mode measurement period.
Hourly or fixed milestones (design, shadow mode, rollout); tell us which you prefer.
If this goes well, we are looking to offer a full-time position to the right person.

To apply, answer these in your proposal

Describe a voi cemail or answering machine detection system you shipped: the approach, the accuracy you measured, and what it got wrong.
How would you resolve the conflict between playing a greeting instantly and needing a few seconds of audio to decide?
Have you worked with Telnyx Call Control directly? Give an example.
Start the first line of your proposal with the word "Telnyx" so we know you read this.

Proposals that are generic or do not answer the questions will not be considered.