Launched this week
LiveReview
Free, Risk-Aware Code Review: Auto Approve Low-Risk Changes
339 followers
Free, Risk-Aware Code Review: Auto Approve Low-Risk Changes
339 followers
Most AI review tools flag issues. LiveReview promises production results: fewer incidents, fewer rollbacks, less senior reviewer load. Every PR is risk-scored. The lowest-risk 30-50% are safely auto-approved; the risky ones go straight to the engineers who should see them. The approach is inspired by Meta's highly successful internal AI review system, where production incidents fell 50x and rollbacks 3x. LLM + incident-trained ML + code-graph analysis. Free and self-hostable.
Free Options
Launch Team
ActVoxBuilt by Humans for Agents
Promoted
In February, you helped git-lrc reach #3 Product of the Day. Thank you.
git-lrc fixed one problem: developers committing AI-written code they hadn't read.
As teams adopted it, a bigger problem showed up at the pull request.
AI made code cheap to write. Reviews stayed expensive.
Every PR waited in the same queue for the same few senior engineers.
A renamed variable and a rewritten payment flow got the same attention.
Seniors became the bottleneck. They skimmed. RISKY changes slipped into production.
Most AI review tools add more comments to the thread.
We measure review by production outcomes: fewer incidents, fewer rollbacks, less senior reviewer time.
LiveReview follows the approach behind Meta's highly successful internal AI review system, RADAR: score every change for risk, fast-track the safe ones, and send the risky ones to the right reviewers.
At Meta, this led to 50× fewer production incidents, 3× fewer rollbacks and 35% faster reviews.
Here's how LiveReview does it:
- Auto-approve: the lowest-risk 30–50% are approved instantly. The rest ALWAYS get a human reviewer.
- Risk-first review: high-risk PRs go to the top of the queue with comments sorted by blast radius, sent to the engineers for review.
- Cognitive Coverage: walk through a PR as slides or a summary, or quiz yourself on it before it merges. Your team still understands the code AI writes.
- Livi Assistant: ask about your engineering data in plain English, on the web or in Slack, Teams and Discord. Track adoption, ROI and the teams that need attention across 50+ actionable graphs.
- Integrations: GitHub, GitLab, Bitbucket, Gitea and Azure DevOps.
- Any AI model: reuse the Claude or OpenAI key you already pay for, or any of 15+ AI providers, or use the LiveReview model.
LiveReview is FREE and SELF-HOSTABLE. Run it on your own infra.
See how it works, then hit Try Free: https://hexmos.com/livereview/
If you like what we're building, star it to help devs discover LiveReview: https://github.com/HexmosTech/LiveReview
@shrsv
noticed you feed production incidents back into the risk score unlike coderabbit.
btw how many merges or rollbacks does the ML need to process before that 30-50% auto approve rate actually becomes reliable for a new team?
@mohsinproduct That's highly org/codebase/deployment specific, but even a few days worth of data can will quickly capture "hotspots" and improve the classification significantly. There are also "high severity" but "rarely changed" things -- which do not become exposed this way. But for most orgs even 2 weeks good enough to capture the landscape I'd say.
@mogabr Hoping to make a contribution to the world of engineering -- including newcomers entering this field!
I'd definitely want to try this. We built our own review setup and are still improving it. I like that LiveReview promises production results, not more comments.
Two things I keep running into. With everyone on our team running 10+ coding agents in parallel, the rebase, re-review, rebase loop burns a lot of time and tokens. And it's hard to give the reviewer enough context to see the whole architecture, not just the diff. @shrsv how does LiveReview handle these?
congrats on the launch shrijith. auto approving low risk prs is where i would be careful. a tiny pricing or config change can look harmless in code but still affect customers. can those be marked as risky too?
@agshin_rajabov Thanks Agshin, for engaging!
Short answer: Yes, the size of the code is only one of the signals we use. We have blast radius analysis based on AST, production feedback, keywords, approval levels and such things to make it more intelligent.
On top of that -- the ML system can learn over time from production behaviors (incidents, rollback) to know which parts of code are tricky, etc as well.
The thing is -- over time -- the system level performance will improve -- lower downtime, reduced incidents, rollbacks, etc. So we aim for that.
By the way you can read Meta's RADAR paper on how this kind of system works in a real world case study: https://arxiv.org/abs/2605.30208
Interesting! So How does it work? Is it like a graphify kind of tool? that looks at blast radius and then does industry standard heuristic bench marks on what it finds etc? I assume no AI since its free and self hostable?
@alchemist_studio Hi Andre, thanks for engaging.
It combined around 20+ "signals" to create risk profile -- including things like graph based blast radius analysis, production performance feedback, code complexity, contributor trust levels, LLM semantic feedback, static analysis, etc. Essentially it's a custom risk scoring formula in a clean workflow.
It combines LLM, traditional ML, static checks -- pretty much any and everything that can contribute to risk profiling (aiming to be comprehensive and providing many levers to customize it exactly to the way an org needs it)
Free and self-hostable with my own Claude key is the part that would actually get me to try it — I don't love routing my code through another hosted reviewer. As a solo dev, the Cognitive Coverage bit is what caught my eye: walking a PR as slides before merging is a nice answer to the 'AI wrote it, I skimmed it' problem.
@jcbuilds We do have a docker image at our Github, you're welcome to try it out. Although we also have a cloud version for quickly understanding how it all works at a high level. And we support pretty much all major providers (you can mix and match) -- so that shouldn't be a problem as well.
Honestly, half the PRs I see barely need a review these days. This makes a lot of sense.
@shricodev1 Absolutely, the trick is in knowing which half is worth skipping :) And even within a PR which parts deserve attention is where we think can make a contribution to your workflow. Thanks for engaging!
before an org has recorded any reverts or incidents, how much weight does the custom ml model get in the risk score?
@bbueno Hi Bruno, thanks for engaging!
When there is no production feedback data, there is no impact from the ML learning component. We go by graph based blast radius analysis, LLM semantic analysis, keyword analysis, code complexity metrics, reachability and many other such signals.
@shrsv makes sense, thanks Shrijith