Loading…
AI Voice Agent Development · Seattle
Voice agents at Seattle scale do not look like a Vapi demo. A marketplace customer-support line at Expedia or Redfin sees hundreds of concurrent calls at peak. An airline IRROPS spike at Alaska Airlines pushes thousands of inbound calls in 20 minutes. An e-commerce post-purchase support line at an Amazon FBA shop has to handle Spanish, English, and Mandarin without a noticeable handoff. The Vapi demo agent is not what survives that traffic shape.
We build voice agents on Amazon Connect, Microsoft Teams Phone, and Azure Communication Services for AWS-native and Microsoft-native teams in the metro. STT through Deepgram Nova-2, LLM through Bedrock Claude Sonnet 5 or Azure OpenAI GPT-5, TTS through ElevenLabs Flash or Cartesia Sonic, all tuned for sub-800ms perceived latency. Observability into CloudWatch or Application Insights with per-call traces.
Every engagement is fixed-scope. Discovery first, fixed quote after, four to ten weeks of build with weekly call demos.
Tell us about the call volume and the stack.
On AWS the architecture is Amazon Connect as the telephony and contact-flow layer, with the agent runtime sitting in a Lambda or an ECS Fargate service called from a contact-flow block. The STT layer is Deepgram Nova-2 streaming for English at 150 to 250ms partial latency. The LLM layer is Bedrock Claude Haiku 4.5 in the same region with streaming enabled for 200-400ms TTFT, falling back to Sonnet 5 for complex conversations. The TTS layer is ElevenLabs Flash v2.5 or Cartesia Sonic for sub-200ms first-audio.
The runtime checkpoints session state in DynamoDB so a Fargate task restart does not drop the call. Every tool call (Salesforce, Zendesk, Snowflake, your CRM) is wrapped in a circuit breaker and an idempotency layer. CloudWatch dashboards track active calls, p50 and p95 TTFT, end-of-turn detection accuracy, and tool-call success rate per thousand calls. SREs debug at 2am from a span trace, not a log search.
On the Microsoft side, the equivalent is Teams Phone Direct Routing or Azure Communication Services Calling as the telephony layer, with the agent on Azure Container Apps in the same region as your Teams tenant. The LLM layer is Azure OpenAI GPT-5 or Claude on Azure (where available) under Microsoft's BAA when PHI is in play. STT and TTS still run through Deepgram, ElevenLabs, or Cartesia, with traffic routed over PrivateLink or VNet integration to avoid public internet.
For scale, both architectures rely on the same idea: bound the number of concurrent sessions per runtime instance so per-call latency stays consistent, and scale out horizontally based on active-call count rather than CPU or memory. Bedrock provisioned throughput or Azure OpenAI PTUs give predictable model capacity with on-demand burst as a fallback.
Six components that show up in every production voice agent we ship for AWS-native and Microsoft-native teams.
Agent attached as a Lambda-backed step inside your Connect contact flow with full access to Contact Attributes. Warm handoff to human queues with transcript carried as attributes. Voice ID for caller verification.
Agent reachable from Teams Direct Routing or Azure Communication Services for Microsoft-stack shops. Integration with Dynamics 365 or ServiceNow for ticketing. Region-local routing for sub-region latency.
Deepgram Nova-2 streaming for English, AssemblyAI Universal for Spanish and other languages. ElevenLabs Flash v2.5 or Cartesia Sonic for first-audio under 200ms. VAD tuning for accurate end-of-turn detection.
Claude Haiku 4.5 or GPT-5 mini for dialog management, with escalation to Sonnet 5 or full GPT-5 for complex turns. Provisioned throughput for predictable capacity, on-demand burst for spikes.
Per-call spans with STT latency, LLM TTFT, TTS first-audio, tool-call latency. Dashboards for p50 and p95 across all stages. Alerts on degraded perceived latency or rising hallucination flags.
Synthetic call replay against labeled scenarios in CI. Task completion rate, response correctness, latency targets, and handoff appropriateness gated before any prompt or model change ships.
Seattle teams ship voice at a scale most demo agents never see. Amazon Connect ships from South Lake Union and runs a sizable chunk of US enterprise contact-center traffic. Microsoft Teams Phone ships from Redmond. Zillow, Expedia, Redfin, and Alaska Airlines each run customer-support operations where peak concurrent calls measure in the thousands. The teams here are not asking whether voice AI works; they are asking how it behaves at p95 during an IRROPS spike or a marketplace incident.
The marketplace shape of the economy matters. Expedia's traveler support lines see volume spikes when a weather event cancels flights. Redfin's home-tour scheduling line sees evening and weekend spikes. An Amazon FBA seller's post-purchase line sees spikes around Prime Day. A voice agent that works at 100 calls a day melts at 5,000 without bounded session counts, idempotent tool calls, and a token-bucket throttle in front of every external system.
We work remotely with Seattle clients on Pacific time. Discovery calls fit between 9am and 5pm PT, weekly demos include live call recordings, and the code ships to a repo you own with full IaC.
Industries where we see strongest fit: e-commerce post-purchase support, travel and airline IRROPS, real-estate scheduling and inquiry handling, B2B SaaS support automation, and Microsoft-shop receptionist and after-hours agents.
Describe the call flow, the cloud, peak concurrency, and which CRM or contact-center stack the agent has to live inside. We'll reply within one business day with an architecture sketch and a price range.