Kevin Lu Profile Banner
Kevin Lu Profile
Kevin Lu

@kevinlu1248

748
Followers
42
Following
23
Media
143
Statuses

Co-founder of Sweep. Prev: MLE at Roblox, NLP Researcher. Try out Sweep at

San Francisco
Joined December 2021
Don't wanna be here? Send us removal request.
Pinned Tweet
@kevinlu1248
Kevin Lu
5 months
In the video below, watch Sweep fix our landing page’s call to action! Sweep is: SOC 2 compliant Fully self-hostable Supports AWS Bedrock 📅 Ready to save your team hours of work each week? Book a demo at our website (thanks again Sweep for fixing that UI bug)! (3/3)
1
0
4
@kevinlu1248
Kevin Lu
5 months
Sweep achieves 15.7% on SWE-bench! Hi everyone, we’re building Sweep, an open-source AI developer that handles the easiest 30% of software tasks. We’re thrilled to announce our results on SWE-Bench! We evaluated Sweep on a random 10% subset of the data. Sweep correctly
Tweet media one
21
23
267
@kevinlu1248
Kevin Lu
10 months
Hi everyone, we’re incredibly excited to launch Sweep V1 today! Over the past few months, we’ve been building @sweep__ai , an AI-powered junior developer. Over the past two weeks, we’ve narrowed our focus and greatly improved: Reliability - generating PRs from prompts
12
14
106
@kevinlu1248
Kevin Lu
5 months
We’re excited to announce that Sweep now uses Claude 3 via both @AnthropicAI and @awscloud Bedrock! Over the last few months, we've used Claude over our customers’ GitHub repositories to fix bugs and ship features. So far, we've seen some pretty great results: - Haiku is
3
7
40
@kevinlu1248
Kevin Lu
1 year
🚀 We just published Sweep's core algorithm, an RAG-based pipeline. Dive into the details at @sweep__ai #code #GenAI #techblog #tech #ai #ArtificialIntelligence
Tweet media one
2
6
41
@kevinlu1248
Kevin Lu
1 year
We just published our blog on how we chunk 2M+ files/day for code search using concrete syntax trees! Would appreciate an upvote and a read at
Tweet media one
2
4
31
@kevinlu1248
Kevin Lu
1 year
Recently, LlamaIndex added Sweep's chunking algo as their default code chunker for retrieval. We just blogged about improving this algo at and you can try out our demo at . #datapipelines #machinelearning #searchenginetechnology
0
7
26
@kevinlu1248
Kevin Lu
1 year
Released our blog on why editing code with GPT-4 is a lot harder writing code from scratch. GPT-4 can’t rewrite long files without writing “# Rest of the code” halfway through or copy code incorrectly, so we work around this at Sweep. More on this at
1
2
21
@kevinlu1248
Kevin Lu
1 year
My friend's a freelance dev that works 24/7 and can multi-task on an infinite number of tasks, with over 1k lines of open-source code contributions, 1.4k GitHub stars and 300M GitHub repos seen. Resume's at and you can hire at
Tweet media one
1
4
20
@kevinlu1248
Kevin Lu
1 year
Thanks @yi_ding for integrating Sweep's chunking algorithm into Llama Index! Excited to see the future of open-source RAG-baser agents!
@jerryjliu0
Jerry Liu
1 year
Text splitting is essential to building a RAG app, but code splitting is under explored - simple delimiters don’t work! The code splitter by @sweep__ai solves this problem: recursive splitting integrated with CST’s of >100 languages 🔥 You can now use in @llama_index 🦙! 👇
Tweet media one
3
21
154
0
4
17
@kevinlu1248
Kevin Lu
1 year
How to get GPT to generate JSON without fine-tuning or functions Just tell GPT to put the answer in XML tags: <answer> { "foo": "bar" } </answer> And parse it with regex: <answer>(.*?)</answer> This helps skip GPT's preamble (like "Sure, let me help you with that...")
3
2
15
@kevinlu1248
Kevin Lu
5 months
About Sweep: We’re building the most capable AI developer. Sweep writes and validates code, then an engineer from your team reviews its work. Saving them hours waiting on tests and builds. We provide on-premise deployments of Sweep, ensuring your code never leaves your VPC.
0
0
13
@kevinlu1248
Kevin Lu
1 year
🎉 Just launched Sweep Sandbox: Sweep guarantees every commit lints and compiles using static code analysis and fixing itself after every file change. Sweep successfully migrated our landing page from React functional to class components () with Sweep map.
2
4
12
@kevinlu1248
Kevin Lu
8 months
In our demo, Sweep improves the is_markdown method to support multiple text files, as well as updating the corresponding tests. To make Sweep easier to adopt, we refactored Sweep to run as locally as possible. Sweep no longer requires a GitHub PAT or GitHub app, and directly
3
4
13
@kevinlu1248
Kevin Lu
7 months
🔒I’m excited to share our guide on deploying Sweep on-premises! Sweep is an AI-powered junior developer that automatically fixes bugs and patches security vulnerabilities. We’re able to locate and make changes in legacy codebases with 20k+ files. This saves developers time by
3
2
12
@kevinlu1248
Kevin Lu
1 year
🎉 More exciting Sweep features 🔴 Sweep live streams progress to indicate which files are done by Sweep (see screenshot) 🔁 Self-review has been drastically improved to minimize mistakes 🐢 Slow mode (using "Sweep (slow):") completes up to 3 self-reviews, ensuring pristine code
Tweet media one
0
2
10
@kevinlu1248
Kevin Lu
5 months
🚨 Has your team’s tech debt been piling up? Sweep is an AI developer that helps your team ship fast. We launched a better planning algorithm last week to help Sweep solve harder issues. This has led to a huge improvement in our internal and customer benchmarks. (1/3)
1
1
13
@kevinlu1248
Kevin Lu
1 year
Check out CodePlan, graph-based LLM code editing: We recently also added graph-based planning for @sweep__ai for Python by linking entities like functions and classes across the codebase. Excited for the future of autonomous code editing agents!
1
2
10
@kevinlu1248
Kevin Lu
10 months
@sweep__ai is currently trending on GitHub for Python Repos! Thank you for all of the support, since our release last week we’ve shipped multiple improvements: - Lexical search is now 1.5x faster for all repositories! - We shipped a new modification algorithm! In our testing,
Tweet media one
2
2
10
@kevinlu1248
Kevin Lu
1 year
Sweep + ChatGPT on Azure OpenAI Integration 🤖☁️ Exciting news — Sweep now supports ChatGPT on Azure OpenAI! Get the power of Sweep with the security and privacy of Azure instances. 🔒 Setup Azure by configuring the following env vars: OPENAI_API_KEY=API KEY
0
2
8
@kevinlu1248
Kevin Lu
1 year
👀Brief preview of a Chrome extension we're building to make creating Sweep issues easier. We have features lined up like file name auto-completion and docs drag-and-drop to make creating detailed issues for Sweep easier @sweep__ai #ArtificialIntelligence #GPT4
0
4
8
@kevinlu1248
Kevin Lu
1 year
We just launched Sweep Map! This breaks larger issues into smaller sub-issues, perfect for repetitive tasks like "Sweep (map): migrate from React class components to function components". See this in action at
0
2
8
@kevinlu1248
Kevin Lu
1 year
🚀Exciting features today for Sweep: 🌐URLs of publicly accessible websites make Sweep read the content 📝Sweep adjusts to issue changes 📜Large files (>600 LOC) get edited 2-3x faster 🔎Search uses a higher-quality embedding model 🎊Sweep built the last three features!
0
3
8
@kevinlu1248
Kevin Lu
1 year
🚀 Just released v0 of Sweep Browser Extension at letting you create Sweep issues from the repo home page, with more features to come like auto-completing file names and docs drag-and-drop. Try it out and give us feedback at !
0
2
8
@kevinlu1248
Kevin Lu
9 months
Yesterday we shipped a fully local instance of Sweep, which uses your Github PAT to create pull requests from Github issues! This calls OpenAI directly from your machine without your code entering our servers. We’re excited for more developers to be able to use Sweep without
1
3
9
@kevinlu1248
Kevin Lu
5 months
“Hey do you have a minute? Can you explain how this function works?” Every software developer has been asked questions like these. There’s an awesome blog by Joel Spolsky where he describes interruptions: “Mutt can’t remember the name of the Unicode version of the strcpy
2
0
7
@kevinlu1248
Kevin Lu
11 months
Sweep's Refactor Revamp! Sweep now refactors each function individually and passes in their dependencies. E.g. if function f calls g, we would pass g's body into the context. We would previously include the entire imported file which can be more granular and confuses GPT-4.
1
0
7
@kevinlu1248
Kevin Lu
1 year
Sweep will now generate a flowchart to show its progress and the rest of its plan more clearly. When the CI/CD checks fail, Sweep shows the newly created tasks to fix these errors. See this example at .
Tweet media one
0
0
6
@kevinlu1248
Kevin Lu
1 year
We just gave Sweep Grep Sweep now searches for specific terms and edits them individually, making repetitive edits like refactors and renames a lot more reliable. E.g. Sweep added time elapsed to our telemetry by searching for all `posthog.capture` calls
0
2
6
@kevinlu1248
Kevin Lu
10 months
Sweep V1: Dashboard for Sweep's Progress You can now track Sweep's progress in real-time on the new dashboard. This feature was created to provide more insight into what Sweep is up to - from search to planning and coding. Take a peek at an example here:
1
0
6
@kevinlu1248
Kevin Lu
2 years
@yvrairport We're still on the plane btw. It's been a good 9 hours. We're almost out of fuel. "We will prioritize this Flair aircraft" my fucking ass. The flight attendants are super helpful and considerate btw.
0
0
5
@kevinlu1248
Kevin Lu
1 year
Sandbox Revamp: We improved the sandbox auto-fix features to use our latest file edit algo. We also rebuilt the UI for sandbox runs, showing every commit & the attached failed sandbox run. See for an example. Expect to see a lot more PRs passing sandbox!
Tweet media one
0
2
5
@kevinlu1248
Kevin Lu
1 year
🎉 Just revamped the Sweep issues page! Now highlighting Sweep's contributions to public repos more clearly! @sweep__ai
Tweet media one
0
2
5
@kevinlu1248
Kevin Lu
1 year
🎊 [BETA] We just updated Sweep to use Elastic License V2 and migrated to Docker to make Sweep deployable on-prem! We rewrote the deployment guide () to reflect this. Excited to see Sweep self-hosted by the community and feel free to ask us questions!
0
1
5
@kevinlu1248
Kevin Lu
1 year
Just Revamped Sweep Sandbox! 🎉 Just gave our AI developer a super linter! Want every commit tested and validated? Sweep runs Trunk ( @trunkio ) to format & check it's changes. Want to use custom dev tools? You can configure Sweep using our setup tool
0
2
4
@kevinlu1248
Kevin Lu
1 year
We're implementing docs search at @sweep__ai (eg. for issues like "Migrate the ReAct agent to LlamaIndex") and we're wondering what frameworks you would like Sweep to index. If there is a library's docs you would like us to index for Sweep to search, feel free to add it at
1
1
4
@kevinlu1248
Kevin Lu
1 year
Last Friday I was asked to share some thoughts on how we're building Sweep and our development of LLM-powered agents. I was thrilled to meet many founders and engineers building exciting tools automating web interfacing. I can't wait for the day when agents can file my taxes or
@AlexReibman
Alex Reibman 🖇️
1 year
We invited 100+ engineers to take an exclusive look at the latest advancements in AI agent ecosystem If you’re not following this space, your job is probably going to be replaced by AI. Here’s what we saw in the AI Agent Developer Arena (🧵):
Tweet media one
Tweet media two
26
110
853
1
0
4
@kevinlu1248
Kevin Lu
1 year
Update: we made it to the front page of Hacker News!
Tweet media one
0
0
4
@kevinlu1248
Kevin Lu
11 months
We just released a free trial for Sweep Pro! Many users aren't sure if Sweep integrates well with their workflow so we released this one-week trial. Check this out at .
0
2
4
@kevinlu1248
Kevin Lu
11 months
We just added a home page to our blogs! We've been covering a lot of content on our technical challenges and ideas recently, so this will make these articles easier to navigate for other engineers building AI agents.
Tweet media one
1
1
4
@kevinlu1248
Kevin Lu
8 months
Sweep can now be installed via NPM, making our installation process much smoother! To download Sweep, check out our installation docs: We also have two major changes since our last release: File creation is now enabled, allowing Sweep to both create and
0
0
4
@kevinlu1248
Kevin Lu
1 year
At a high level, we run a similar algorithm to LangChain's recursive text splitter on a concrete syntax tree (CST) parsed using Tree Sitter. The low-level details are in the blog post.
1
0
4
@kevinlu1248
Kevin Lu
11 months
@sweep__ai now auto-imports undefined variables in Python using the library importmagic. GPT-4 would often forget to import modules it needs to use, such as when writing unit tests, so this greatly improves the reliability of Sweep. Check out Sweep at !
0
0
3
@kevinlu1248
Kevin Lu
10 months
@RyanEndacott @wwzeng1 @sweep__ai Some examples we found recently that worked really well are adding type hints and adding a new route to your website. For example, Sweep created the initial about us page and GitHub banner on our landing page.
0
0
3
@kevinlu1248
Kevin Lu
5 months
⏰ We also sped up Sweep by 3x, reducing our average time to initial PR from 15 minutes → 5 minutes! ✅Sweep now uses this extra time to test and validate your code for up to an hour, ensuring the generated code is review-ready and passes your CI! Rather than just writing the
1
0
3
@kevinlu1248
Kevin Lu
5 months
@Leoagua1 We're using Claude Opus mainly.
0
0
2
@kevinlu1248
Kevin Lu
1 year
☁️🛠 Sweep AI Cloud Deployment Guide V2 🛠☁️ We revamped Sweep's self-hosted deployment guide to add instructions for deploying on AWS, GCP, or Azure! 🌩️ 🔗 Learn More: #CloudDeployment #AWS #GCP #Azure #SweepAI 🌟
0
1
2
@kevinlu1248
Kevin Lu
4 months
AI isn’t going to replace software engineering, but it can make us better. As an engineer, you might have 30 minutes between meetings to review a PR and move your team forward. Slow code review bottlenecks engineering teams as they wait on teammates, and speeding through proper
2
0
5
@kevinlu1248
Kevin Lu
1 year
🚀 Our blog on Sweep's core algo is on the Hacker News front page! You can find us at  and Ctrl-F: Using Retrieval Augmented Generation to clear our GitHub Backlog Would love to hear your thought!
Tweet media one
0
1
2
@kevinlu1248
Kevin Lu
10 months
@handfuloflight @sweep__ai Typescript is supported! We support all languages but have some modules like unit testing and refactors built vertical for Python.
0
0
2
@kevinlu1248
Kevin Lu
1 year
What if you could leave your tech debt to an AI? 🤖 Three ways people handle tech debt: 1. before you release 2. after you release 3. or just forget about it Use @sweep__ai , and forget about your tech debt ✨. Check this out at
Tweet media one
0
0
2
@kevinlu1248
Kevin Lu
11 months
Looks like GitHub is having an incident so Sweep may not be working as intended
0
0
2
@kevinlu1248
Kevin Lu
1 year
@dhaynes23 @sweep__ai That's fine, we'll let you know when the file modification improvements are out. Thanks for the feedback!
1
0
2
@kevinlu1248
Kevin Lu
1 year
Pay by Usage: you can now pay for Sweep to handle a single GitHub issue using GPT-4. You can access this at by entering your GitHub username. We added this since we're actively improving Sweep for GPT4 rather than 3.5 and wanted more people to use GPT4.
0
0
2
@kevinlu1248
Kevin Lu
1 year
🚀 Been a blast having Sweep used by the Llama Index team!
@yi_ding
Yi Ding -- prod/acc
1 year
Version 0.0.14 of @llama_index TS is out! 🚀 Huge shout out to @sweep__ai and @y_am_a_da for the contributions. New and improved 🦙 Llama2 🦙 integration. Full changelog here:
1
7
18
1
0
2
@kevinlu1248
Kevin Lu
1 year
@yi_ding We just use XML at Sweep and you rarely run into unparsable outputs in the first place. Spends a few extra tokens but saves a lot of headache.
1
0
1
@kevinlu1248
Kevin Lu
8 months
Please reach out if you’re interested in booking an onboarding session here: You can also install Sweep Assistant here: Thanks for all of the support, and we're excited to keep shipping 😀
0
1
1
@kevinlu1248
Kevin Lu
1 year
We just open-sourced Sweep, a Github assistant that writes PRs for tickets!
@sweep__ai
SweepAI
1 year
Check out our repo at
0
0
0
2
0
1
@kevinlu1248
Kevin Lu
1 year
What happens if an AI-powered junior dev uses a pre-configured all-in-one devtool like @trunkio ? We wrote an article on how we get Sweep to use dev tools like Trunk, the orchestration and infra complexities, and how that makes Sweep a lot more powerful
0
0
1
@kevinlu1248
Kevin Lu
2 years
Customer service first said the flight was only delayed for 1 hour, forgetting to account for the tarmac delay. After a long conversation, they agreed to recompensate me. They then sent another email saying again they only delayed for 1 hour, as if the conversation never occurred
Tweet media one
Tweet media two
1
0
1
@kevinlu1248
Kevin Lu
1 year
@maxhodak_ Hey Max we're building exactly this at . Sweep makes bug fixes and implements small features by creating PRs and validates them with configurable linters and tests.
0
0
1
@kevinlu1248
Kevin Lu
1 year
@FZaslavskiy @WizardLM_AI There's speculation that the current GPT-4 is poisoned by having humaneval in the training data so it's not a fair comparison.
0
0
0
@kevinlu1248
Kevin Lu
1 year
@silennai @maxhodak_ Thanks for the ping!
0
0
1
@kevinlu1248
Kevin Lu
1 year
Feel the same way. Can't wait to make a Llama 2 doppel of my co-founder
@halvarflake
Halvar Flake
1 year
every time I read the @modal_labs docs I have 25 random ideas and a deep yearning for more free time to play
0
1
15
0
0
1
@kevinlu1248
Kevin Lu
2 years
Is Flair's customer service usually this negligent? I can't tell if these are bots since even bots now can adequately address an issue like this.
0
0
1