Coval Runs 10,000 Call Failures. Its New Hire Must Ship Them.

Four Roles, One Loop

Coval, a voice-AI evaluation platform that closed a $28 million Series A in June 2026 led by Norwest with participation from Base10 Partners, Twilio Ventures, and Y Combinator — bringing total capital raised to $31 million since its 2024 launch, is hiring for four roles that expose a widening gap in the AI stack: companies can build voice agents faster than they can prove those agents work in production. The four openings — Deployment Strategist, Forward Deployed Engineer, Founding Product Engineer, and Account Executive, map directly to the bottlenecks that appear when evaluation moves from research notebooks into customer CI/CD pipelines.

The Deployment Strategist role, posted days ago, sits at the intersection of the platform's simulation engine and the customer's production reality. Coval's platform automates scenario generation for voice and chat agents, runs continuous evaluation to detect regressions, and supports sentiment analysis and performance monitoring. A strategist who can translate "we simulated ten thousand calls and found a regression in handoff logic" into a customer's deployment timeline is the connective tissue between the product's CI/CD evaluations and the revenue that follows.

The Forward Deployed Engineer and Founding Product Engineer, both listed for six months, signal that the core product surface is still expanding. The platform's value proposition — "simulate thousands of scenarios engineers don't have to manually test", requires engineers who can build the simulation infrastructure itself: scenario generation, reproducible speech-provider benchmarks, human-review loops, and the integrations that slot into existing development workflows. The "founding" prefix on the product role indicates the codebase is early enough that architectural decisions still compound.

The Account Executive, with a band reaching $350k, reflects a go-to-market motion that sells to technical buyers at the frontier of voice AI. Coval works with "companies at the frontier of voice AI, from Fortune 500s to the fastest-moving startups," per its own description. That buyer knows the difference between a demo that passes and an agent that holds up under production traffic. The sales hire needs enough technical fluency to credibly discuss regression detection, sentiment drift, and the specific failure modes of voice pipelines — latency, interruption handling, provider-specific quirks, without handing off to an engineer for every question.

What the Work Actually Demands

Coval's product scope makes its hiring priorities legible: the company builds the evaluation layer that sits on top of every voice agent stack, and the people who join need to have wrestled with the same problems Coval's customers face. That means hands-on experience designing and running evaluation pipelines — not just consuming them, carries more weight than a degree from a top program or a brand-name internship.

The platform's architecture reveals the competency stack. Coval's simulation engine requires engineers who understand how to model stochastic voice interactions at scale: generating realistic personas that interrupt, hesitate, switch languages, and call from noisy environments across 27 voices, 10 languages, and 20 acoustic environments. Candidates who have built or maintained similar simulation harnesses — especially those who have calibrated metrics against human judgment and wired the results into CI/CD, map directly to the work. The onboarding flow that Coval ships to customers (Setup → Connect Agent → Discover Use Case → Build Test Cases → Select Metrics → Launch and Review) is essentially a checklist of the skills the team exercises daily: translating a business use case into a test set, selecting metrics that capture resolution, safety, and experience, and closing the loop so human review sharpens the next simulation run.

Voice-specific domain knowledge is non-negotiable because the failure modes live in the audio layer. The platform measures time-to-first-audio, latency variation, and word-error-rate across transcription and TTS providers; it surfaces telephony artifacts — SIP headers, jitter, packet loss, that text-only evals miss. Engineers who have debugged a production voice agent know why VAD models with limited context break turn-taking, why full-duplex speech-to-speech models invalidate turn-based eval harnesses, and why sarcasm and sentiment disappear when the LLM only sees transcripts. Founder Brooke Hopkins has said publicly that in 2026 "you should probably not be writing your own prompts… you should probably like have an eval and get a model to write your prompts and just like hill climb." That mindset, treating prompt optimization as a search problem grounded in evaluation data, is exactly the muscle Coval hires for.

Cross-functional fluency matters because Coval's buyers span engineering, product, QA, operations, and compliance. The platform integrates natively with Langfuse, LangSmith, Arize, and Datadog via OpenTelemetry, and customers expect the evaluation layer to slot into existing tracing infrastructure without a rewrite. A candidate who has instrumented a voice agent with structured spans, correlated simulation results with production metrics, and presented findings to a compliance reviewer speaks the language of every stakeholder in the room.

Portfolio evidence beats credentials. A GitHub repo that implements a reproducible speech-provider benchmark, public methodology, code, and input data, signals more than a master's thesis on LLM theory. A write-up of a simulation campaign that caught a regression before launch, or a dashboard that reduced manual QA by an order of magnitude, demonstrates the loop Coval sells: simulate, observe, review, improve. The company's own benchmarks and cookbooks are built by the team; contributors who can extend that library with new verticals (healthcare intake, debt collection, IT helpdesk) or new failure modes (policy traps, tool errors, multi-language switching) add immediate leverage.

The pedigree that matters is the scar tissue from shipping evaluation pipelines that caught real failures in production.

Why Voice Breaks the Chatbot Playbook

Voice AI evaluation is not chatbot evaluation with a microphone attached. The distinction shapes everything Coval builds, and everything it hires for. A voice agent runs three models in parallel: one transcribes speech, another reasons through a reply, a third speaks it back. That architecture mirrors the perception, planning, and control stack of a self-driving car, a comparison Hopkins draws from her years leading evaluation infrastructure at Waymo. Both systems operate in continuous time, both must react to unpredictable environments, and both fail in ways that static test suites never catch.

The failure modes are distinct to voice. An agent that sounds flawless in a demo trips over accents, talks over background noise, and freezes when a caller goes off script. Evaluation has to judge timing, turn-taking, interruptions, audio artifacts, tool calls, and caller emotion, not just the final transcript. Coval's platform is purpose-built for this: audio processing and quality analysis, telephony latency measurement, transcription error analysis, and evaluation of agent workflows across customer service, sales, financial services, and healthcare. The company runs tens of millions of simulated tests on a single voice agent, and its customers, over 60 organizations including Zoom and Deepgram, report cutting manual QA by up to 30x while deploying agents up to 10x faster.

That scale changes what "domain expertise" means. It is not enough to have fine-tuned an LLM or built a RAG pipeline. Candidates who understand how voice agents actually fail, how a 200-millisecond latency spike breaks turn-taking, how background music confuses a transcription model, how a policy trap in a healthcare call requires a different escalation path than a sales objection, bring signal that a resume full of generic "AI engineering" bullets cannot. Deepgram's COO Anoop Dawar said voice agents introduce a new level of complexity compared to traditional software testing. Zoom's head of CX AI product, Ram Rajagopalan, called reliability and observability a top priority as voice AI moves into customer-facing production. Both companies partnered with Coval because they have lived the failure modes.

The market context amplifies the premium on this knowledge. More than $7 billion flowed into voice AI in the first quarter of 2026 alone, with forecasts passing $20 billion by 2031. Startups like Bland have raised tens of millions to build the agents themselves. Twilio's voice-AI revenue has been climbing fast, and Twilio Ventures backed Coval's Series A rather than building its own testing tool, a signal that even the infrastructure layer wants independent evaluation. Over 100 agent startups emerged from Y Combinator's three 2024 cohorts. One, /dev/agents, raised a $55 million seed at a $500 million valuation less than a year after founding. Every one of those companies will need to prove their agents work before an enterprise puts them in front of a customer.

Coval's simulation-first method, transplanted from autonomous vehicles, only works if the simulations reflect the real physics of a phone call: jitter, packet loss, overlapping speech, the caller who mumbles a credit-card number while a dog barks. Building those simulations, calibrating the metrics, and integrating the tracing and review queues is what Coval's solutions-engineering and product teams do daily. The company plans to expand those teams with the Series A proceeds, alongside deeper simulation, more integrations, and stronger human-review tooling. Candidates who have wrestled with telephony stacks, built evaluation harnesses for real-time audio, or shipped voice agents that survived production traffic will move through Coval's screen faster than those who have only read about the problem.

The Deployment Strategist role posted last week will close when Coval finds someone who can translate ten thousand simulated failures into a customer's go-live date. That translation, from simulation to production, is the loop the four roles are built to close.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.