Available for focused production sprints

Rokas Remeika

Production Realtime Voice & Telephony Engineer

I build and harden production realtime voice systems, from LiveKit Agents and WebRTC through SIP/PSTN, with observability, fault reproduction, and automated evaluation to validate the work.

View specialist sprints
  • LiveKit Agents SDK · Python
  • LiveKit SFU · WebRTC · TURN
  • SIP/PSTN · realtime voice AI
01 / agent AgentSession Python agent runtime
02 / media LiveKit SFU WebRTC and TURN
03 / telephony SIP gateway signaling and RTP
04 / network PSTN real telephone calls

Production proof

~2,700 Recorded end-to-end voice-agent runs
SDK internals Production failures traced and fixed below the application layer
Carrier / PSTN SIP signaling, RTP/media, routing, and failure handling
Full realtime stack SFU · SIP · TURN · Redis · TLS · deployment and recovery

The system works. Production is less forgiving.

I work across the agent, media, infrastructure, and telephony boundaries where intermittent realtime failures are hardest to reproduce and easiest to misdiagnose.

01

Dropped or silent calls

Separate application stalls from signaling, RTP, WebRTC, TURN, SIP, carrier, and provider failures, then engineer reconnect, watchdog, and mid-session recovery.

02

Latency and turn-taking drift

Trace timing across STT, LLM, TTS, AgentSession, media transport, interruption, and barge-in behavior.

03

Changes that regress real calls

Use real RTC scenarios, fault injection, recordings, and predefined acceptance criteria to show whether a change improved the system or moved the failure.

04

Retry and recovery failures

Harden dial idempotency, call duration limits, failure classification, admission control, and deployment recovery before retries amplify the problem.

Production proof / anonymized system

One engineer across the whole call path.

I built and run a production AI interview system using the LiveKit Agents SDK in Python, a self-hosted LiveKit media server, and a SIP gateway that places real outbound PSTN calls. It has handled hundreds of production AI phone calls across 60+ languages.

The evaluation system joins independently through livekit.rtc as a real RTC peer, sends real audio, runs autonomous scenarios, and captures recordings, traces, scores, and outcomes as black-box evidence.

Across roughly 2,700 recorded runs, unattended batches combine deterministic and LLM-assisted scoring, adjudication, build-identity checks, and fault injection. Predefined acceptance criteria have rejected apparently good changes when they introduced different regressions.

That work has caught calls that connected into an empty room and a silent mid-interview freeze with no crash or useful error. When application-level changes were insufficient, I traced the failure into SDK and plugin internals, and fixed it below the application layer. The resulting fix was covered against recurrence.

Runtime LiveKit Agents SDK · Python AgentSession in a production worker
Infrastructure Full realtime stack SFU · SIP · TURN · Redis · TLS · deployment and recovery
Evaluation ~2,700 recorded RTC runs Real audio, batch execution, scoring, adjudication, and build verification

Focused engagements. Concrete engineering artifacts.

Fixed time and deliverables, scoped around the stack I can inspect and change.

Regression and evaluation are part of how I validate production work, using recorded calls and predefined acceptance criteria. If you specifically need a reusable real-call test harness, I can scope that as a standalone engagement.

01 / realtime voice Focused sprint

Production Realtime Voice Sprint

Build or harden the realtime voice layer, from LiveKit Agents and WebRTC through STT, LLM, TTS, turn-taking, tools, and production observability.

  • LiveKit Agents / AgentSession implementation
  • STT → LLM → TTS realtime pipeline
  • Latency, turn-taking, interruption, and barge-in
  • Tool calling, state, lifecycle, and reconnect behavior
  • WebRTC/media debugging and production observability
  • Regression coverage needed to validate the work
From EUR €6,500
02 / telephony Expanded scope

Realtime Voice + SIP/PSTN Sprint

Take the realtime voice system through to real phone calls, adding the SIP, PSTN, carrier, routing, media, and production behavior that telephony introduces.

  • Everything relevant from the realtime voice layer
  • LiveKit SIP, SIP trunks, and carrier integration
  • Inbound / outbound PSTN calling and routing
  • DTMF, transfers, and human handoff
  • SIP signaling and RTP/media interoperability
  • Real telephone validation and regression scenarios
From EUR €9,000

Best fit

For teams where voice has real operational weight.

CTOs and founderswith production calls and a hard reliability problem.

Agencies and AI consultanciesthat need white-label depth on the realtime or telephony layer.

Engineering leadsshipping voice as a feature without a dedicated media or SIP specialist.

Portrait of Rokas Remeika

Founder-led by design.

You work directly with me from traces and architecture through implementation and handoff. Dreamchasers is the Lithuanian company I operate and contract through.