How Docket uses Entelligence to make sure the same bug never ships twice
The Reviewer That Doesn't Forget
How Docket uses Entelligence to make sure the same bug never ships twice
Docket AI builds the AI teammate that turns scattered GTM knowledge into instant answers for revenue teams. Their agents live inside Slack and Teams, fielding live sales and technical questions where a wrong answer reaches a rep, or a customer, in seconds.
That same bar for reliability is what they hold their codebase to. So the most valuable thing Entelligence does for Docket isn't catching a bug. It's remembering one, and making sure it never ships again.
One incident, learned once, blocked everywhere
Most code review tools can't carry a lesson from one incident forward into every review that follows. Entelligence can.
A single PR on Docket's codebase established a rule: authorization state has to be revoked before the underlying record is deleted, or you end up with dangling access. Entelligence captured that as "Learning 21."
Weeks later, on a completely different repo the original PR never touched, the same failure mode showed up again. A MongoDB record was set to be deleted before its permission tuple was revoked. Entelligence flagged it and cited Learning 21 by name.
That's not a linter running the same static rule twice. It's a reviewer that remembers what already went wrong, so the same mistake doesn't reach production on someone else's repo, weeks later, without a human having to reconnect the dots.
Docket now has 37 org-specific guidelines like this one, all learned from their own PR history, all actively reused. The longer Entelligence runs, the more of Docket's hard-won lessons it carries into every review.
The catches that become rules
Each finding below is exactly the kind of thing that, uncaught once, becomes the reason a rule like Learning 21 gets written. Entelligence's job is to catch it the first time, then never let it repeat.
Findings | |
|---|---|
A live JWT leaking over an unverified connection. | A config file forwarded a caller's bearer token to an internal service with TLS verification turned off, exposing it to MITM interception. Flagged, fixed. |
A suspended user reactivation bypass. | An admin invite flow quietly reactivated suspended users and cleared their session invalidation flag before they'd even accepted the invite. A security gate with a hole in it. |
A vulnerable dependency floor missing where it mattered. | A known CVE fix was applied in sub-packages but skipped at the root, leaving the main app still exposed. |
An unverified supply chain download. | A CI runner pulled a GitHub Actions tarball without checking its SHA-256, while still running an EOL Node.js version on infrastructure with private key and registry write access. |
A NOT NULL migration that would have broken production. | A new column shipped with a default that only existed in Python, not the database. It would have broken every existing row on write. Flagged twice on the same PR before it merged. |
Fifteen months a customer, the last three months in numbers
Docket activated Entelligence in April 2025. It's our longest tenured customer, still running today, 15 months in, across all 68 active repos spanning their AI agent platform, backend services, infra, and desktop and browser clients.
We don't lead with vanity metrics. Here's what the last three months alone looked like:
5,273 review executions at a 97.7% success rate
5,764 inline comments left across the codebase
2,514 of those verified as hits, addressed by a developer or auto-resolved
470 distinct PRs touched by a verified fix
~3.2 minutes average review time
37 org-specific guidelines learned from Docket's own PR history
That's a 43.6% hit rate over the last three months, across 68 repos. Not a pilot spike that fades by month three, but a reviewer that has kept getting more specific to Docket's codebase for 15 months and counting.
What the data does and doesn't prove
Security scans on 108 of Docket's PRs turned up zero SonarQube flagged vulnerabilities. The real catches above came from Entelligence's own review, not a bolted-on scanner.
8 PRs were reverted across the full 15 months as a customer, out of 964+ reviewed in the last three months alone.
What we will say: 43.6% of everything Entelligence flagged changed the code, and the reviewer doing the flagging gets sharper with every incident it sees. Docket didn't run a trial. They made Entelligence the part of their process that doesn't forget.
Fifteen months in, Entelligence.ai isn't the tool that catches Docket's bugs. It's the reason the same one never comes back.