Cekura - The self-improvement loop for voice agents
by•
Cekura is the testing, observability, and self-improvement platform for production voice and chat AI agents. It simulates thousands of scenarios, catches failures, diagnoses the root cause, rewrites prompts and config, then re-validates with a full regression sweep. Unlike tools that hand failures back to your team, Cekura closes the loop by fixing the agent itself and proving the fix holds without overfitting.


Replies
Hey Product Hunt!
Sidhant here, co-founder of Cekura
Today we are launching self-improving loops for voice agents.
Fixing a voice agent has always been fragmented. Your testing tool tells you what failed, you diagnose it from transcripts, patch the prompt, re-run, and something else breaks. The tools find problems, but the fixing has always been a human walking between them.
Cekura collapses that loop. It runs thousands of simulated calls and groups every failure, explained in plain English. Optimise agent hands them to the Cekura Agent, or any coding agent you use, Claude Code, Codex, anything. It reproduces each failure, makes the change, and reruns until everything passes, then verifies nothing else broke. You review the diff.
Two rules: it must reproduce a bug before fixing it, and every fix is proven on simulated calls on a clone, never your live agent.
Free for everyone to try, starting today. We are in the comments all day. If you want to chat more, please feel free to book time here
Root cause diagnosis for voice agents is genuinely hard because failures compound across turns. If it traces that accurately, that alone is worth the pitch.
@peter_victor Completely agree. We trace the full conversation to find where behavior first diverged, rather than patching only the final symptom.
How does Vocera handle accuracy and context when conversations get more complex or users have different speaking styles? Congrats @kabra_sidhant & team!
@hamza_afzal_butt Thank you! Cekura tests agents across varied scenarios, speaking styles, interruptions, and multi-turn context—not just isolated turns. It also preserves conversation history and evaluates whether the agent completed the intended outcome, so accuracy holds up as interactions get more complex.
@hamza_afzal_butt Our evaluators are written against intent, not exact wording — conditions like "the agent asks for your name" match semantically regardless of how a specific agent phrases it, and grading runs over the full conversation using LLM/rubric-based judging rather than pattern matching. So the same test generalizes across agents with different tones, styles, or conversational complexity, without needing to be rewritten per agent.
@satiwick1 Thanks for your encouraging comment. Means the world to us. If you get a chance do try cekura for your agent at cekura.ai
The rule that a bug has to be reproduced before it gets fixed, and validated on a clone instead of the live agent, is the part that stands out to me. I build a voice companion that calls aging parents daily, and my worst fear is a silent regression in an emotionally sensitive call that nobody catches until it actually matters. How do you generate realistic edge cases for messy real speech (long pauses, hearing loss, someone talking over the agent) versus clean synthetic voices? That gap is where most of our failures hide.
@igorgurovich That's exactly the gap we designed for. Our simulations use Conditional Actions to reproduce real pauses, mid-sentence interruptions, and real-voice recordings, not just clean TTS , along with personalities for speaking pace and background noise. That's where messy-speech edge cases get caught before production.
The rule that a fix has to be reproduced and validated on a clone before touching the live agent is the right engineering call for anything running in production customer support. The part I want to understand is the CI/CD integration: when a new prompt version or agent config gets committed, does Cekura run the regression suite automatically as a pipeline step that can block a deploy, or is the test run still something the team triggers manually? That boundary between an automated gate and a manual check is usually where teams cut corners under deadline pressure.
@hazy0 Cekura evals can run as a step in your CI/CD pipeline via our GitHub Actions integration. We also provide a curated Infrastructure Suite that we recommend adding to that pipeline for infra regression checks.
the loop-closing part is the interesting claim. the failure mode i would worry about is the simulator quietly becoming the thing you optimise against, so the agent gets very good at passing your scenario set and no better in production.
we run agents against third-party systems we do not control, and the split that ended up mattering most was "the agent did the right thing" versus "the other side actually accepted it". those two look identical in a transcript and come apart constantly in practice.
curious how the regression sweep decides a fix generalised rather than fit the scenarios. is the held-out set drawn from real production calls, or generated the same way the training scenarios were?
@whateverneveranywhere That’s exactly the failure mode we guard against. We separate “the agent attempted the right action” from “the external system accepted it” using provider state, tool results, and outcome signals—not transcripts alone.
The regression sweep includes the full validation set plus happy/edge cases, with overfitting checks for scenario-specific fixes. Production-derived held-out cases are the strongest validation.
It's like living in the future
I'm a huge fan of the work the Cekura team is doing. Closing the development loop with testing, evaluation, and paths to verifiable improvement is the single most valuable thing you can do in an agentic coding context. This is hard for voice agent development, for a bunch of reasons, including that there are many moving parts, gathering metrics robustly can be tricky and subtle, and success criteria are complex. Cekura has created great building blocks and high-leverage agent skills/workflows for solving these problems. Congratulations to the team on this launch, and all the effort and good thinking that went into it!
@kwindla Thank you so much! We really appreciate your help with this
the without overfitting part is doing a lot of heavy lifting in that description, and its the right thing to be worried about. an auto-fixer that cant prove the fix generalized is just prompt roulette.
@alex_watson2110 Agreed , that's exactly the risk we designed against. Every applied edit passes through an Overfitting Gate before validation, then gets re-run against the full evaluator set, so a fix only counts once it holds broadly, not just on the case it was written for.
@dileepchagam late reply, sorry. that gate answers the per-fix version of the problem well.
the one it does not cover is the cumulative version: if the fixer keeps optimising against the same evaluator set, that set quietly becomes the training set, and you end up fitting the benchmark rather than the behaviour. worth holding a slice of evaluators back that the fixer never gets to see, and only checking them periodically. if the visible score and the held-back score start drifting apart, that is your tell.