Customer discovery
Public research becomes a testable hypothesis: which intents, surfaces, latency ceilings, and escalation rules make voice useful rather than decorative?
01 · A potential prospect thesis
Magalu is a focused customer hypothesis, not a generic demo.
The large Brazilian retail and tech platform already has Lu1, a massive Brazilian consumer surface4, and a WhatsApp customer journey where users naturally send voice notes3. My proposal tests whether ElevenAgents can become the voice specialist inside that customer workflow.
Voice deployment
PT-BR voice judgment, brand-voice constraints, sample curation, and a future Professional Voice Clone (PVC) path are treated as deployment work, not just creative polish.
Technical bridge
The repo includes Python tools, mock APIs, logging redaction, tests, and an orchestration boundary for handoff and tool calls.
Commercial frame
The pilot is packaged around CFO-readable value, licensing critical path, risk gates, and a Day-60 decision point.
02 · Demo
Talk to Lu, or listen without microphone access.
The public demo is intentionally narrow: a first feel for voice tone, narration-before-action, PII-safe phrasing, and human handoff. The production plan lives in the repo.
live · ElevenAgentsBETA
Ask about delivery, return, order status, or something out of scope.
Hosted with an ElevenAgents widget5. Luiza is the Phase 1 persona, rendered on a premade Multilingual v2 voice as the bridge until licensed Professional Voice Clone is in place. PT-BR primary, configured around three behaviors: explain before acting, mask sensitive data, and hand off clearly when the task leaves scope.
Try in PT-BR
“Oi Lu, tudo bem?”
“Quero falar com uma pessoa.”
What the repo proves: a tested local runtime with mock Magalu tools, dynamic tool responses, narration-before-action, formatted CPF/CNPJ/phone/email redaction at the logging layer, and deployment docs for architecture, risk, quality, prompts, and rollout.
03 · Deployment model
Voice specialist for Phase 1. The role expands as the data earns it.
The strongest first-phase integration story registers ElevenAgents as the voice specialist beside Magalu’s existing customer, cart, search, and support agents. Phases 2 to 4 expand the surface area as the data and the customer’s confidence earn each next step.
“ElevenAgents plugs in as the voice specialist inside Magalu’s existing orchestrator1, then returns structured context when the turn belongs to another agent or a human.”
Architecturecustomer surface → orchestrator → voice runtime
CollaborationMagalu Eng ↔ Strategist ↔ ElevenLabs Research/Product
What changes for the customer
Voice notes can get voice replies for in-scope intents. Out-of-scope turns are handed to a human with the transcript, reason, and next step intact.
What I would measure
Containment, CSAT delta vs text-only baseline, p95 first-audio latency, escalation precision, and MOS for Lu’s voice.
What I would feed back
Brazilian CPF/CNPJ speech handling, BR retail voice presets, timeout heuristics, and retry-with-idempotency patterns for future deployments.
04 · Rollout and commercial frame
A two-week spike, a Day-60 review, then scale only if the evidence is clean.
From prospecting through close to implementation: the path a Magalu engagement actually runs.
Parallel critical path
Talent licensing
PVC rights open
Cast or close
Production voice live
Discovery and integration spike
- Discovery CX leader + engineer interviews, intent segmentation, voice suitability, PII and legal, latency budget.
- Spike One intent, one retry path, one structured handoff, mock tool bridge, logging redaction, baseline p50/p95.
Commercial review
Three deal structures on the table:
- Hybrid floor-plus-upside Recommended. Defended floor near monthly spend + 15-25% share of cart-recovery lift.
- Pure outcome-based Linked to cart-recovery lift. Cleanest CFO story; ElevenLabs carries ramp risk.
- Implementation fee + license Procurement-conservative fallback.
Phase 1 review
- Metrics Conversion lift · CSAT delta · MOS for Lu’s voice.
- Gate decision Expand to Phase 2, redesign, or sunset.
Licensed voice path
- Trigger Day-90 metrics pass.
- Transition Luiza bridge voice → licensed PVC.
- Governance Rights scope, expiration windows, version control, review.
Pass gates
- Containment above the Day-0 customer-defended floor (illustrative: ≥ 60% for in-scope intents).
- MOS ≥ 4.0 / 5.0 for Phase 1 voice.
- p95 first-audio under agreed threshold.
- Escalation precision accepted by human agents.
Stop gates
- Two consecutive weeks below containment floor.
- CSAT below the text-only baseline.
- PII/logging issue in replay or monitoring.
- Voice-brand risk not cleared with legal/talent owners.
05 · About the candidate
Customer ownership + commercial instinct + voice craft.
Sixteen years across customer-facing release operations, commercial negotiation, voice and localization, and QA leadership at studios including Keywords, Riot, Altagram, and Maximal.
The commercial instinct comes from co-founding Neutrimem: prospecting at Gamescom, structuring commercial terms, driving engagements to close, and owning delivery through platform-holder certification.
The combination I would bring to ElevenLabs Brazil: customer discovery, voice judgment, deployment-grade prototyping, deployment risk discipline, and the ability to translate between business, engineering, and product teams.