Now: Building in AI (DM me to ask about it) | Prev Founder of https://xitter.cf/t.co/JGaSpIzy2N (YC, a16z)

Joined April 2016
Yesterday, we emerged out of stealth to announce Jo, an open source, local, private, AI assistant. You can get work done with your voice, from anywhere. Even while doing yoga. I've used Jo to get work done during yoga sessions countless times.
2
1
5
408
Today, we're announcing Jo. Now, you can get your work done through voice. Jo an open source, local, private, AI assistant. Jo clears your inbox, schedules meetings, and handles follow-ups, all while your phone stays in your pocket. I started using Jo when I stopped enjoying fundraising. Zoom calls are awkward. Where do you look? How do you smile? How do you prevent yourself from looking at you? One day, my friend suggested I do fundraising calls over the phone, and I never looked back. That's where the idea for Jo came about. A tool where you can get your work done with voice. I've used Jo to do work on a run, on a jog, commuting, doing yoga, cooking, folding laundry, and every situation in between. And yes, it's local, secure, and private. We don't see any of your data or the tool calls themselves. And if you select local-first models, none of the data leaves your phone. We spent the last couple of months building and refining, testing with users and iterating on Jo, and we would love you to try it out. We are nowhere near perfect, but we hope we're good enough, fast enough, local enough, and private enough to spark some inspiration. You can either download the app, or you can directly load it onto your phone. Both are linked below. We've added a skill file inside of our repo to let Claude or ChatGPT know how best to install it. Try it out, push it to its limits, and let me know what you think.
30
13
2
153
14,554
We're running our AI assistant's brain on an iPhone. It needs a small AI model for listening and picking the right tool or action to use. We tested several models on 16 common tasks. Apple’s built-in phone model got 9 of 16 right. Google’s open model Gemma, as-is, got 11. Gemma with a small add-on we trained for Jo got all 16. (That add-on is called a LoRA: a thin layer of extra training on top of a model.) Impressive work from Google, with a little help from us.
3
7
665
Whenever a user gives feedback on our AI voice assistant, they do it through the assistant. We've built it so we get feedback and an additional test of the assistant's capabilities.
1
266
This is what you need to do while Cursor spawns 1,500 sub agents.
1
2
296
Agent engineering setups are so 1:1 and personal. When we interview engineering candidates, we don't ask them to change their setup live. Instead, the candidate brings their own AI coding setup instead of using ours. They spend the hour building against a real project spec. Before any of this happens, we send a one-page document on how we use AI tools at Kiloforge. That way, we're not spending time explaining what we expect, or setting up an environment.
1
1
4
252
More startups should pick CEO based on the CEO chair.
3
257
We've done a lot of work on our on-device AI assistant, making sure as many API calls as possible become cached storage. This became all unread mail, and the last seven days of your inbox. All on-device.
1
157
We just deleted almost all of our personal assistant's backend, and moved that logic on-device. A personal AI assistant that's privacy-first is what Apple should have built. We're doing it right.
3
1
4
295
If you're wondering what the Kiloforge team does while Claude is working...
2
239
We are TikTok dancing to PMF
1
197
I'm more worried now than I was a month ago about model step-changes. Voice model regression testing could be a real startup opportunity. We switched from GPT Realtime to GPT Live and four pieces of tuning broke at once. Now, we're forced to build aggressive regression tests so we know what breaks when models change. You can't even diff voice output. It's like text model release problems on steroids.
2
1
4
356
If you need help from an investor, make their job as easy as possible. One of our investors emailed a list to the whole portfolio. It named people affected by layoffs at their companies. Instead of replying "We're hiring, send anyone who fits" I replied and specifically asked for two people. Now, it's one job instead of homework on their end. I try to think about most interactions this way.
2
309
10x cheaper tokens mean I can serve power users of our product in a completely different way. Our token-heavy users used to worry me. If costs fall 10–100x while models stay this powerful, they stop being a cost problem. It opens up iterations of @KiloforgeAI we didn't think were possible. We're all lucky to have this much competition amongst model providers.
1
119
Kiloforge is hiring engineers, and the job is in person in SF, five days a week. When we get intros for candidates, we always lead with this. It's a dealbreaker for some, and both parties need to learn dealbreakers as early as possible. It's cheaper to learn over first email than after 2+ rounds of interviews.
2
1
1
11
898
This is your sign from the @kiloforgeai team to touch grass
1
121
At 8am, our user logged a bug. By 2pm, the fix is a pull request, and that night it's deployed. If you told me this was the reality 10 years ago, I would have thought you were crazy.
1
1
140
To avoid useless PRs, our coding agents have a "Not Verified" section at the end that details what the agent didn't check. One recent PR named three gaps: - It was never run on a real device. - It skipped the team's validation check. - One data field had no test of its own after being merged into a bigger bundle. Now, anyone reading the PR months later sees all the shortcuts.
1
1
2
278