▮▮ EDGAR CORDERO

AI Solutions Architect · Voice Systems

Voice agents that sound human — and ship like software.

I design and run production telephone AI for companies that can't afford a bad first ring. Conversation design, real-time pipelines, telephony, and the ops behind them — from first call audit to an agent your customers thank by name.

▶  Hear an agent work Book an intro call
620 ms
median voice-to-voice latency
140k+
production calls handled
88%
calls resolved without a human
24/7
coverage — every line, every night

Recorded call

Listen in on a call.

A real call handled end-to-end by the agent. The transcript below is synced to the recording — click any line to jump there.

Inbound · chipped tooth, booked for next morning
Northgate Dental — after-hours line · real recording
00:00

Press play to listen in — the transcript follows the recording as it plays. Click any line to jump to that moment.

Try it live — call Jarvis

+1 (917) 451-6400
A live demo agent on the same system as the call above — different voice, same craft. No signup, no sales pitch — just a phone call, 24/7.
  • Pick a scenario. Jarvis walks you through three: book a dental appointment, grill a restaurant host about menu and hours, or get order help from customer support.
  • Interrupt it. Talk over it mid-sentence, change your mind, switch scenarios halfway through. Adapting gracefully is the hard part.
  • Ask it anything. Including whether it's a real person — it'll tell you the truth. Nothing you book or ask is real, so push it as hard as you like.

Selected work

Agents in production, not in demos.

Every engagement below is live today, measured weekly, and owned end-to-end.

Inbound · Healthcare

Appointment scheduling for a 9-location dental group

Replaced an overflow answering service with a voice agent that books, reschedules, and confirms directly into the practice management system — including insurance pre-checks.

AGENT · “I have Thursday at 9:10 or next Tuesday at 2:40 with Dr. Okafor — Thursday's the sooner one for a cleaning. Want me to lock it in?”
1,400/mo
calls handled
78%
booking completion
610 ms
median latency
PipecatTwilioDeepgramElevenLabsClaude
Outbound · Solar

Lead qualification for a regional solar installer

Calls web leads within 60 seconds of form submission, qualifies roof type, ownership, and utility bill, then books the survey — before a competitor's SDR has opened the CRM.

AGENT · “Got it — you own the home and the bill runs about $240 a month. That's squarely in the range where panels pay for themselves. Can I get a site surveyor out Saturday morning?”
58 s
median speed-to-lead
3.1×
more surveys booked
41%
connect rate
PipecatTwilioCartesiaHubSpot API
Inbound · Property management

After-hours maintenance triage across 4,000 units

Distinguishes a burst pipe from a burnt-out bulb at 2 a.m. Emergencies page the on-call tech with a structured summary; everything else becomes a ticketed work order by morning.

AGENT · “Water coming through the ceiling counts as an emergency — I'm paging the on-call technician now and texting you their ETA. Is the unit's water shut-off accessible to you?”
100%
answer rate
-72%
non-urgent callouts
4.8/5
tenant CSAT
PipecatSIP trunkDeepgramAppFolio API
Platform · Cost architecture

Voice-stack cost rebuild for a high-volume operator

Re-architected a per-minute vendor stack into a self-hosted pipeline: benchmarked open TTS models against ElevenLabs, added caching for repeated prompts, and moved inference to on-demand GPU capacity.

RESULT · Same perceived voice quality in blind listening tests. The invoice, not the experience, is what changed.
-22%
cost per call minute
-15%
end-to-end latency
6 wks
to full migration
Kokoro TTSAWS EC2vLLMTerraform

Under the hood

The architecture behind the voice.

Three views of the same system: the path a syllable travels, the AWS footprint it runs on, and the loop that keeps it honest.

CALLER PSTN · mobile TWILIO media streams · WS audio AGENT RUNTIME · PIPECAT ON ECS FARGATE TURN-TAKING VAD · barge-in interruption recovery STT Deepgram streaming partials LLM Claude streamed · tool use TTS ElevenLabs / Kokoro sentence-chunked partials ~150 ms first token ~300 ms first byte ~120 ms TOOL CALLS PMS · CRM · calendars — run while speaking synthesized audio streamed back — playback starts before the sentence is finished TOTAL VOICE-TO-VOICE ~620 MS MEDIAN — EVERY STAGE STREAMS; NOTHING WAITS FOR A FULL RESPONSE
The real-time path. Latency is a budget, not an outcome — each stage gets an allocation and is measured against it in production. The pipeline streams end-to-end: STT emits partials while the caller is mid-sentence, the LLM starts speaking on its first token, and tool calls run in parallel so booking a slot never produces dead air.
CALLER PSTN TWILIO SIP · media AWS · SINGLE REGION VPC public subnet NLB TLS · WebSocket private subnets ECS FARGATE agent workers scales on live calls EC2 GPU self-hosted TTS on-demand · $0 idle RDS POSTGRES call state transcripts bi-directional audio TTS stream STT · LLM · TTS APIS VIA NAT EXTERNAL PROVIDERS DEEPGRAM streaming STT CLAUDE API reasoning · tools ELEVENLABS TTS fallback CRM · PMS bookings land here LAMBDA webhooks · evals S3 recordings SECRETS MGR key rotation CLOUDWATCH latency alarms
The deployment. Agent workers autoscale on concurrent calls, so a Monday-morning spike and a quiet Sunday cost what they should. GPU capacity for self-hosted TTS is on-demand — it exists only while call volume justifies it — and every provider on the right is a swappable module behind a provider flag, which is how the 22% cost rebuild in the case study above was possible without downtime.
LIVE CALLS production traffic Twilio · all lines · 24/7 RECORDINGS + TRANSCRIPTS S3 · every call captured stereo audio + turn-level log EVAL SUITE nightly scenario replays graded against real conversations CANARY DEPLOY shadow first, then ramp 5% → 100% · instant rollback FIX + TUNE prompt diffs · pipeline changes new edge cases become new tests SCORECARD containment · latency · task success week-over-week trend every call nightly batch graded runs regressions changes redeploy REAL-TIME GUARDRAIL CloudWatch latency + error alarms page a human in minutes, not mornings THE LOOP RUNS EVERY NIGHT — AN EDGE CASE A CALLER FINDS TODAY IS A PERMANENT TEST BY MORNING
The feedback loop. Shipping is the midpoint, not the finish line. Every night the eval suite replays real conversations against the current agent and grades containment, latency, and task success; a regression blocks the next deploy, and anything a caller surprises the agent with becomes a permanent test. Changes return to production through a canary ramp with instant rollback. The dashed path is the fast lane: when latency or error rates blow their budget, a human gets paged in real time — the nightly loop is for the slow drift no alarm can see.

Capabilities

The whole pipeline, one pair of hands.

Vendor-neutral by design — every layer below is swappable, benchmarked, and chosen per project.

Orchestration

Pipecat, LiveKit Agents, custom Python pipelines. Turn-taking, barge-in handling, interruption recovery, and tool calls that fire mid-sentence without breaking the flow.

Telephony

Twilio, SIP trunking, number porting. Inbound IVR replacement, outbound campaigns with compliant calling windows, warm transfer to humans with full context.

Speech

Deepgram & Whisper for ears; ElevenLabs, Cartesia & open models for the voice. Blind-tested voice selection and per-use-case latency/cost tuning.

Reasoning

Claude and GPT-class models, prompt architectures for reliability under interruption, structured tool use into CRMs, ERPs, financial and scheduling systems.

Infrastructure

AWS (ECS, EC2, Lambda), Terraform, on-demand GPU. Agents that scale to call spikes and cost nearly nothing at 3 a.m. when the lines are quiet.

Quality & ops

Eval suites on real transcripts, latency budgets per turn, call analytics dashboards. Weekly regression runs so the agent that shipped is the agent still running.

How an engagement runs

Four weeks to a live line.

WK 1

Call audit

I listen to your real calls — recordings, dispositions, drop-offs — and map which conversations an agent should own and which it must hand off.

WK 2

Conversation design

Scripted happy paths are where voice agents die. I design for interruptions, accents, background noise, and the caller who changes their mind twice.

WK 3

Build & integrate

Pipeline, telephony, and the integrations that make it useful — calendars, CRMs, ticketing. Shadow mode first: the agent listens before it speaks.

WK 4

Launch & operate

Gradual traffic ramp with human fallback, then ongoing evals, latency budgets, and monthly reporting. I run what I build.

Contact

Your phone is ringing. Let's answer it.

Tell me about the calls you're missing — I'll come to the intro call with a point of view on whether a voice agent is actually the right fix. Sometimes it isn't; I'll say so. And because I trained in business and accounting before AI, your CRM, ERP, and billing workflows aren't a foreign language to me.

Edgar.Cordero@Computer.org Replay the demo
BaseCiudad de México, MX · remote worldwide
Responsewithin one business day
Live demo+1 (917) 451-6400 — Jarvis · 24/7
Credentials AWS Certified Machine Learning — Specialty · M.S. Computer Science, AI/ML · Member, IEEE & IEEE Computer Society MLS-C01 · active through Aug 2028 · validation no. c0576d4eb3e945848bfbdb095efcc31f verify: aws.amazon.com/verification · Credly badge · IEEE member no. 102572019
LanguagesEspañol (native) · English (bilingual)
Elsewherelinkedin.com/in/edgar-cordero-ai