coval

Building a code reviewer that engineers actually trust is much harder than it looks. Our engineer Dana wrote up the real story behind ours, including the parts that didn't work.

Early versions ran on frontier models and hallucinated defects that weren't there, at $2.07 per review. The fix wasn't a bigger model, but a smaller, more deliberate system: a bounded read-only agent, an ensemble of cheaper models that de-correlate better than four runs of one strong model, and a simple constraint requiring every claim to quote the exact changed line.

That last change alone took false positives from 38% down to 98% verified accuracy.

The result: a reviewer that catches real defects at eleven cents per review in production.

Full breakdown from Dana here: coval.ai/blog/how-we-bu…

Most TTS latency benchmarks stop the clock the moment any audio arrives, silent or not. We changed ours to count that silence. The day the update shipped, @gradiumai reported latency moved from 172 ms to 430 ms. No model change or infrastructure change. Their audio was arriving fast, it just opened with up to 655 ms of dead air before anything audible.

Gradium used that signal to retrain their default model. Now they're hitting 214 ms, about 170 ms faster.

If you're picking a TTS provider on a latency number, make sure to check what that number is actually measuring.

Read more in our latest blog post with Gradium: coval.ai/blog/time-to-f…

There’s a persistent belief in startups that you have to choose between moving fast and building stable systems.

Early on, it feels true. Infrastructure looks like overhead. Monitoring, evals, rollback logic, routing discipline — none of it directly ships a new feature. When you’re racing to find product-market fit, stability can feel like something you earn later.

But in systems like voice AI, the tradeoff turns out to be mostly imaginary.

What actually slows teams down isn’t infrastructure. It’s the absence of it. It’s spending days debugging a regression because no one captured per-turn latency. It’s pushing a prompt update and not knowing whether it improved resolution or just shifted failures somewhere else. It’s being afraid to deploy because you can’t predict the blast radius.

When you build observability and structured evaluation early, releases get less dramatic. You know what changed. You know what broke. You can attribute performance movement to a specific component instead of debating guesses in Slack. Rollbacks become procedural instead of emotional.

Speed without feedback is chaotic. Stability without iteration is stagnation. The teams that compound fastest figure out how to build both at the same time — tight feedback loops, clear instrumentation, and the confidence to ship because measurement is built in.

Over time, the supposed tradeoff disappears.

What remains is a system that improves quickly precisely because it’s stable enough to trust.

🎉 Exciting News! Please join us in welcoming Rob Young to the Coval team!

Before joining Coval, Rob designed products at Google and Apple, and founded and led the product design studio OM. He brings a wealth of experience crafting elegant, user-centered experiences at some of the world's most innovative companies.

Outside of work, you'll find Rob spending time with his family, skateboarding, doing yoga, and painting.

Welcome to the team, Rob! 🚀

🧵 This week in conversational AI:

🎤 @OpenAI Realtime takes a big step forward — sub-500ms latency, sharper function calling, non-verbal cue detection, seamless language switching, and SIP (telephony) support. Already benchmarked at coval.ai, and the gains on instruction following are a real unlock for production-grade voice agents.

🧪 @cartesia demos custom evals for voice agents — defining, validating, and reusing LLM-powered metrics on real calls. A strong signal that observability is becoming core infrastructure for scaled agents.

🔗 @DeepgramAI × Cloudflare : Deepgram’s STT + TTS now run directly on Cloudflare’s global edge. Low-latency voice agents deployed serverlessly — making conversational AI faster and more accessible worldwide.

💡 Together AI named an IA40 Winner in the “Enabler” category — powering over 800K developers building AI-native apps. Great to see Cartesia , Arcee AI , and ElevenLabs also recognized, showing how fast the real-time voice ecosystem is maturing.

📣 Coming up (Sept 3, SF): 𝗩𝗼𝗶𝗰𝗲 𝗔𝗜 𝗣𝗿𝗼𝗻𝘂𝗻𝗰𝗶𝗮𝘁𝗶𝗼𝗻 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲𝘀, hosted by @Covaldev , @rimelabs, and @ConverseNowAI. We’ll dig into why models stumble on words like Chipotle, tradeoffs between emotional range and consistency, scaling TTS in production, and war stories from the trenches. If you’re building voice AI, this is for you.
P.S: We'll be sharing more behind-the-scenes of our Coval Benchmarks 2.0, with a special focus on TTS! Make sure to sign up for the event! luma.com/dd7lgjbc

What did I miss?