How smallest.ai Keeps 5,000 Businesses' Calls From Going to Dead Air
The Reviewer Watching the Line Before the Call Drops
How smallest.ai keeps a bad deploy from becoming a dropped call
smallest.ai builds real-time voice AI infrastructure for the enterprise. Their platform spans Atoms, a real-time voice-agent platform that plugs into business systems for support, lead qualification, and outbound calls, and Waves, its text-to-speech and voice-cloning engine spanning 30+ languages. Founded in Bengaluru in 2023 by Sudarshan Kamath and Akshat Mandloi, former engineers at Robert Bosch, and backed by Sierra Ventures, smallest.ai's voice agents have already handled over 1 million calls for more than 5,000 businesses across India and the US.
Their codebase spans atoms-platform, waves-platform, lightning-asr-offline, pipecat, cookbook, and normalization_service_v2: a low-latency ASR/TTS pipeline built for sub-second response times, a telephony and voice-agent orchestration layer routing live calls in production, and billing and credit systems metering usage across enterprise accounts. In voice infra, a regression doesn't show up as a stack trace somewhere in a log. It shows up as a dropped call, or a customer billed for a conversation that never happened. That's the bar.
Reviewing the pipeline that can't afford to hang
A team shipping voice infra at this pace can't manually comb every diff for the concurrency bug, the silent regression, or the crash-on-deploy that sneaks in when everyone's moving fast. Something has to catch it before merge, every time, without becoming the thing that slows the team down.
Over a four-week window, Entelligence reviewed roughly 128 PRs across smallest.ai's six active repositories, leaving close to 396 total review comments, of which Entelligence authored around 352 (88.9%). That makes it the team's dominant automated reviewer by a wide margin. No other review bots run on these repos; human reviewers and GitHub's own advanced-security bot account for the rest.
The catches
Roughly 32 confirmed developer fixes were committed directly after an Entelligence code review comment, with response times as fast as 3.5 minutes. Half of all confirmed fixes landed within five minutes of the comment posting, which is less "we'll get to it" and more the team treating the bot's findings as blocking. A few PRs got closed without merge, minutes after Entelligence flagged something serious enough that patching wasn't worth it.
Findings | |
|---|---|
A concurrency clamp that could never unstick. | atoms-platform #2004: once available concurrency went negative, a validation bug meant users could never reduce usage to escape the over-limit state. Fixed in 7.5 minutes, and the fix commit mirrored Entelligence's own wording almost word for word. |
An assistant stuck listening after the response ended. | cookbook #6: a removed early-exit condition meant the receive loop only ended on timeout or error, silently adding up to 30 seconds of dead air to every completed response, plus a related bug blocking reconnection entirely. Fixed in 3.5 minutes. |
Billing bugs that broke reporting and locked out free-tier users. | waves-platform #632: removing a credits reset broke the analytics contract and corrupted usage reporting, while a separate credit guard blocked free-tier users from using the product at all. Fixed in about 2 hours. |
Language auto-detection silently disabled. | pipecat #477: a dropped default forced every request into English-only behavior. |
A guaranteed crash from a removed import. | pipecat #475: the import was gone but still referenced elsewhere in the same file, guaranteeing a runtime crash the moment that code path ran. Fixed in 6 minutes. |
Two telephony bugs, one PR. | pipecat #472: a sample-rate mismatch that would break telephony audio, plus a dropped call-end handler that left agent-initiated calls open indefinitely. |
A payment processing bug misassociating invoices. | atoms-platform #2036: a non-deterministic database update was silently misassociating Stripe invoice IDs whenever two purchases shared the same amount. Fixed in 24 minutes. |
An unhandled promise rejection that could crash the whole process. | atoms-platform #2027: a fire-and-forget async block with no error handling, capable of taking down the entire Node.js process. Fixed in about 2 hours. |
A PR abandoned rather than patched. | pipecat #471: Entelligence flagged three separate runtime-breaking validation issues, and the author closed the PR two minutes later instead of fixing around them. |
What that adds up to
Runtime crashes prevented: guaranteed NameError/AttributeError/ReferenceError bugs caught across pipecat, atoms-platform, and lightning-asr-offline before they reached main
Data integrity protected: billing and analytics corruption, misassociated Stripe invoices, and a negative-concurrency bug, all caught pre-merge
Production availability protected: a process-killing unhandled rejection, a 30-second user-facing hang, and a URL-normalization regression, all caught pre-merge
Breaking changes caught: a silently reverted billing behavior and a call-termination regression, both flagged before shipping
The numbers
It's a smaller window than some of our longer-tenured customers, four weeks across six repos, but the pattern is already clear:
Entelligence authored 88.9% of all review comments left on these repos, 352 of 396 total, with zero competing review bots picking up the rest
Roughly 1 in 4 PRs reviewed (32 of 128) drew a confirmed developer fix committed straight after an Entelligence comment
Half of those confirmed fixes landed within 5 minutes of the comment posting, the fastest in just 3.5 minutes
At least one PR was abandoned rather than patched, closed without merge minutes after Entelligence flagged it as not worth saving
Closing line
Four weeks, six repos, roughly 352 comments authored, zero competing review tools in the mix, and dozens of confirmed fixes, including a payment misassociation bug, a process-killing unhandled rejection, and a PR abandoned two minutes after Entelligence flagged it as not worth patching. Commit messages routinely echoed Entelligence's own language back, which is about as clear a signal as a team can give that it trusts what the bot is telling it.