Gradon (YC F26)

Make your AEC firm AI-native.

[Gradon (YC F26) @heygradon]

@yousefhll Planes

[Yousef @yousefhll]

Vendo lets your users create: plane games????

With Vendo, users don't just create simple UIs and graphs, they can generate whole apps that do real things.

Here's how we made it work:

> The agent writes a real React/TSX screen. We compile it, scan its imports and tool use, type-check it against the host’s actual schemas, run its queries against real data, and reject invalid trees and dead controls before it paints.

> The code runs as Preact inside a sealed QuickJS WebAssembly VM. No ambient DOM. No network. No clock. It can only emit a data-only UI tree made from Vendo primitives and components the host registered.

> Generated code cannot call your API directly. When someone clicks a button, QuickJS sends the tool name and arguments across the VM boundary as plain data. The host executes that call as the signed-in user, through the same auth and approval system as every other Vendo action, then sends the result back. QuickJS resumes the original handler with its local state intact.

> If the app needs a backend, Vendo can give it a persistent sandbox with allowlisted network access. Host tool calls still come back through the same guarded boundary.

> The host renders everything using its own theme and registered components. It looks native because it is built from the product’s actual design system.

That is how we get from “generate a widget” to “let users create arbitrary apps inside the product.”

All open source. Repos below.

[Gradon (YC F26) @heygradon]

The leading tool for generating product UI.

[Yousef @yousefhll]

Announcing GenBench: an open benchmark for generated product UI.

A polished screen is not necessarily a working product. It can miss the request, misuse a tool, wire controls to nothing, or display numbers the underlying data does not support. We built GenBench to measure those failures.

Vendo produced 99% working screens with 100% data integrity. It beat the best agent in 1/3 the time and at 1/7 the cost.

[Gradon (YC F26) @heygradon]

@yousefhll @nourzahzah let's go

[Nour Zahzah @nourzahzah]

A big thing we focused on with Vendo Knowledge Bases was making sure the agent never mixes up what it actually knows with what it's guessing at.

No company wants a customer getting a confident answer that turns out to be the agent filling in a gap with something plausible-sounding instead of real.

A few things that make that true:

> Every answer sourced from a knowledge base carries a visible source tag. If there's no tag, the agent isn't claiming it as fact.
> Knowledge bases are pre-built per industry, not assembled on the fly from whatever the agent happens to already know.
> Every query against a knowledge base gets logged, so you can see exactly what it pulled and from where, not just what it said.

More on how the industry packs actually get built in a longer writeup soon.

All open source.

[Yousef @yousefhll]

A big thing we focused on with Vendo was making sure that all the data it displays is 100% accurate. No enterprise wants their agent displaying hallucinated data to their users.

We used a bunch of tricks to make this happen:
> Any numbers/or data displayed in the apps the Vendo Agent creates can ONLY be derived from a API call

> We have a judge that reads every screen. And we give it a calculator (see: ability to write simple JS code) to go in and make sure that if the app tells the user a number is something, it is actually that thing (ie. if the app says it is showing a sum, it actually computed the sum)

> Very detailed input and output schemas for every single tool. The more details you give the agent, the better.

A bunch more stuff too (detailed write-up coming soon).

Benchmarks releasing tomorrow, showing just how much better we are compared to everything else out there. All code is open-source in the repo below!

[Gradon (YC F26) @heygradon]

@yousefhll nice

[Yousef @yousefhll]

A big thing we focused on with Vendo Guard was making sure every action is provable after the fact, not just approved before it.

No enterprise wants to take our word for what an agent did with their customer's data.

A few things that make that true:

> Every tool call passes through Guard before it runs. No exceptions.
> Low risk actions execute on their own. Anything that actually matters requires approval first, before it happens.
> Every single action gets logged, not just the ones that needed approval.

More on how we score risk per tool in a longer writeup soon.

All open source.

[Gradon (YC F26) @heygradon]

@abhinavops yes

[Gradon (YC F26) @heygradon]

@mal_shaik Original

[Gradon (YC F26) @heygradon]

@danielxfarrell TAKEOFFF

[Gradon (YC F26) @heygradon]

@doudou_octopus awesome

[Gradon (YC F26) @heygradon]

@zachdotai things are changing

[Gradon (YC F26) @heygradon]

@zakyhassan super next level

[Gradon (YC F26) @heygradon]

@nourzahzah awesome

[Gradon (YC F26) @heygradon]

We did a thing.

[Yousef @yousefhll]

We learned how to fly a plane to tell SF one thing.

Introducing the Vendo (YC S26) Full-Stack Agentic Interface. A production-ready agent for your users with everything they need.

It takes real actions for your users. And it can even build whole new features, micro-apps, and automations, right on top of your product.

Built-in permissioning, tool guards, user context, and more. Everything you need for a user-facing agent.

One command to install. Open source.

@runvendo @nourzahzah

[Gradon (YC F26) @heygradon]

@yousefhll @ycombinator @nourzahzah @garrytan Use Vendo :)

[Gradon (YC F26) @heygradon]

@yousefhll cool

[Gradon (YC F26) @heygradon]

@itsmarinos 🫡

[Gradon (YC F26) @heygradon]

@samelliottsam yay