Ryan Prasad

Accepted Karpathy’s Agentic Engineering job interview challenge:
production Twitter clone for agents, AI simulates activity, survive AI red team 😱

[Ryan Prasad @ryanprasad_ai]

https://t.co/pbBHRKh4Yy

[Ryan Prasad @ryanprasad_ai]

@VaDOT Working on a budgeting AI to help NOVA commuters with toll costs and want to use SmarterRoads pricing data.

Is a bulk download of historical data available? Don't want to risk trouble by scraping.

Thanks for making this open to the public. It's a great service.

[Ryan Prasad @ryanprasad_ai]

@petergyang It's about paying for fast mode and prepping the agent for overnight work.

[Ryan Prasad @ryanprasad_ai]

Grappling with security implications of agent having write permissions on a fresh personal email account.

I like YOLO, but now this is making me a bit uncomfortable.

[Ryan Prasad @ryanprasad_ai]

@forgebitz Hot take: specializing is a trap, especially when the niche is small

[Ryan Prasad @ryanprasad_ai]

Vendor lock in detected.

Hermes is suggesting CloudFlare or Porkbun as the new domain registrar. AWS Route53 has a $40 delta.

Leaning towards paying the convenience fee...

[Ryan Prasad @ryanprasad_ai]

@nicbarkeragain I experience reversion to the mean.

[Ryan Prasad @ryanprasad_ai]

@Teknium @jphorism Thanks for the tip! May give this a shot. Was considering migrating to google

[Ashwin Gopinath @ashwingop]

https://t.co/VdueqcyzBm

[Ryan Prasad @ryanprasad_ai]

Gave up on getting Strix to work with the Codex bridge (so as to use subscription,) but ran into tool calling issues that others have documented.

Switching to Sonnet via API. Not ideal, but I don't think we have a choice right now.

[Ryan Prasad @ryanprasad_ai]

Our budget auto enthusiast AIs had some stuff to say to the pentester! 🤣

[Ryan Prasad @ryanprasad_ai]

Pentest bot signed up and authed. Wonder what it is up to now...

[Ryan Prasad @ryanprasad_ai]

The only thing more fun than building something is blowing it up. Never done a destructive pentest using an AI tool like this. Let's see how badly I mucked it up.

[Ryan Prasad @ryanprasad_ai]

@grok

[Ryan Prasad @ryanprasad_ai]

Looks like the challenge may be over a bit early due to guardrails using Strix:

"API call failed after 3 retries: This content was flagged for possible cybersecurity risk... To get authorized for security work, join the TrustedAccess for Cyber program: https://t.co/VgSJ297yK4"

[Ryan Prasad @ryanprasad_ai]

@OpenClawTips Used Grok to scope out the spec for the (not) Twitter clone project.

In hindsight, it actually may have been the right model for the job. Knew exactly what features to implement and which ones to steer away from.

[Ryan Prasad @ryanprasad_ai]

Quick demo of Karpathy's Agentic Engineering job interview challenge demo. (Not) a Twitter clone with AI users. Front end, back end, database, AWS Kubernetes deployment.

5 days. No code.

Needed to demo before pentest and find out where I screwed the pooch.

Article tomorrow.

[Ryan Prasad @ryanprasad_ai]

Okay! We are closing in on the end of the challenge. Site is deployed live on AWS EKS, and we have bots talking about old cars.

https://t.co/eIGgR2Bcty if you want to check it out.

Gotta start documenting before pentest tomorrow.

Very doable week-long take home interview.

[Ryan Prasad @ryanprasad_ai]

We be using Tailscale to get the bots hooked up and posting while keeping the door shut to X trolls.

Also useful for pentesting!