Sharath Challa

Founder, V12 Labs. 10+ yrs shipping software. Now building AI agents, RAG and workflow automation that ask a human before anything risky. Notes from real builds

[Sharath Challa @srcbuild]

We named our growth engine Vyra.

The prompts stopped being the interesting part a while ago. The graph is.

Campaign → audience → discovery → research in parallel → evidence → qualify → enrich → draft → review → send → learn

Not enough evidence? It loops back and digs more. Nothing goes out until it passes review.

[Sharath Challa @srcbuild]

@finnbuilt 12yrs😅

[Sharath Challa @srcbuild]

@hustle_fred list existing demands and find gaps in them

[Sharath Challa @srcbuild]

Some github repos can save you weeks of work.

What’s one repo you discovered that made a huge difference to your project?

Let’s uncover some hidden gems.

[Sharath Challa @srcbuild]

@Dusyantkmr first and last

[Sharath Challa @srcbuild]

Most AI outbound tools start with a list of leads.

I'm building Vyra(AI GTM engine) to start with the why.

I just started the local run, and it already found 259 market observations → 3 ICPs → 7 accounts under review.

Research agents find signals and evidence before any account is qualified.

What would you want the AI to verify first?

[Sharath Challa @srcbuild]

@axhiii88547 both

[Sharath Challa @srcbuild]

@skpnky Yes

[Sharath Challa @srcbuild]

Anthropic’s Frontier Red Team published research investigating whether AI models could eventually replicate across ordinary internet-connected devices.

https://t.co/kqqAhhyvcE

[Sharath Challa @srcbuild]

One detail in Anthropic’s AI-worm research stuck with me: they ran a 26B model across 15 $55 Android TV boxes.

Slowly (0.39 tokens/sec), but it ran.

What interests me isn’t the speed. It’s how we’d contain an agent that no longer depends on a central server.

[Sharath Challa @srcbuild]

Was looking at LangChain’s new ScaleDown integration for Deep Agents.

Their docs show a tool result going from 18,402 to 3,110 characters by keeping what’s relevant to the agent’s question.

I like that the agent can still fetch the full result if it needs more.

[Sharath Challa @srcbuild]

The hidden risk in agent memory isn't forgetting. It's remembering something that is no longer true.

Researchers at ByteDance and Scale AI tested memory-management approaches across 5 LLMs and found that ExpeL/ACE-style memory rules produced 7–13.5× more stale answers than a verified approach.

This challenges the assumption that more persistent memory automatically makes agents better, while giving AI engineers a concrete architectural principle they can apply to long-running systems.

My take: memory needs revalidation, not just retrieval.

[Sharath Challa @srcbuild]

@malakhovdm mostly how we deliver. we build AI-powered solutions for clients, but also use AI heavily in our own workflows.

[Sharath Challa @srcbuild]

AI made building easier. Not selling.

Building an AI-native services company and learning that firsthand.

Founders, what actually worked for getting your first customers?

[Sharath Challa @srcbuild]

@KiranTarafder True, this was a tested hypothesis.
dhh introduced exactly this idea under a project called ONCE.

[Sharath Challa @srcbuild]

@NickpxJ @bryan_johnson and you share same line

[Sharath Challa @srcbuild]

@MakadiaHarsh with the same conviction I started agency

[Sharath Challa @srcbuild]

@Nate_Esparza unable to sell

[Sharath Challa @srcbuild]

@solopribuilds AI can write better code than top engineers but it needs architecture steering and proper skills to verify end to end