Building Reliable Voice Production Agents: Why Standard Benchmarks Miss Production Failures
Event Live Stream Link : https://go.foundercoho.com/livestream-8-11
📝 Event Description AI voice models are getting scary good, but sounding human in a 30-second demo is only half the story.
We don't ship code straight from a raw LLM without tests, harnesses, and safety guardrails. Yet voice has no equivalent layer today, so the audio you ship sounds perfect in review—and then mangles a customer's name, a drug, a tracking number, or a date on the one line that matters most, and a deal quietly dies.
At this live FounderCoHo session, we skip the high-level hype for an honest, tactical tour of the "Last 5% Problem" in voice AI. We’ll explore where these failures cluster across English, Chinese, French, and Spanish, why standard benchmarks miss them completely, and the routing-and-evaluation layer voice production is still missing, closing with a live demo.
Whether you are building voice agents, AI customer service, or multimodal LLM products, you will leave with a clear playbook on how to test, red-team, and trust the AI voice you ship.
🎙️ Featured Speaker:
• SOOHYUN BAE — Founder & CEO, Podonos [YC W22 / Georgia Tech ML PhD]
Soohyun is the Founder & CEO of Podonos, an automated performance and analysis platform for AI voice models powered by AI and 150,000+ global evaluators. A Y Combinator alumnus (W22) with a PhD in Machine Learning from Georgia Tech, Soohyun was previously the Co-Founder & CTO of Bobidi (community-based AI model red-teaming) and AR Mapping Engineering Lead at Niantic, where he helped build planet-scale 3D mapping systems.
• Moderated by JING CONAN WANG — Founder & CEO, DeepVista AI, Co-Founder of FounderCoHo
🛠️ What we'll cover:
• The Failure Map: Where AI voice models break down in production across English, Chinese, French, and Spanish—and why leaderboards miss them.
• The Missing Layer: Why voice needs its own testing, routing, and evaluation harness before hitting production.
• Real-World Red-Teaming: How to systematically stress-test voice models for edge cases, accents, and high-stakes domain vocabulary.
• Live Demo: A live walk-through showing real-time voice failure modes and how to resolve them before your customers notice.
🎟️ Registration: Virtual Livestream. Please RSVP to receive the streaming link.
Meet your Host! FounderCoHo is a community and media platform for founders to share their hard-earned wisdom and connect with fellow founders on their journey.
With 195,000+ subscribers across YouTube and Substack, we've featured industry leaders, including Lyft's Co-Founder, the inventor of CRISPR, and AlphaChip's co-author, alongside many successful founders and senior executives. We host monthly in-person founder meetups where members exchange insights and build meaningful connections.
Explore more: https://foundercoho.com/
Where to find us: