Krishna

21 | Building in AI | Founding AI/ML Engineer | Super30 - @100xdevs

[Krishna @KrishXCodes]

2 years ago I had no real experience.

Today my first full-time job is
Founding AI/ML Engineer at a US startup.

No fancy background.
Just consistency and a lot of building.

Here’s the real story 👇

[Krishna @KrishXCodes]

I used to think shipping fast meant being sloppy

Turns out shipping fast just means to put something imperfect in front of real people before you feel ready

You’re never gonna feel ready. So now is the timing.

[Krishna @KrishXCodes]

@AnthropicAI Curious how much of this is focused on capability discovery of the model vs deployment safety

[Krishna @KrishXCodes]

This feels like a big shift for real codebase work 🧐

The bottleneck is less "can the model write code?" and more "can it plan, split and coordinate the work better?"

[ClaudeDevs @ClaudeDevs]

New in Claude Code (research preview): dynamic workflows.

Claude writes an orchestration script on the fly, then spins up a large fleet of coordinated subagents in parallel to take on your most complex tasks.

Use the word "workflow" in a prompt to get started.

[Krishna @KrishXCodes]

@builtat2am This is a good reminder that projects are not just resume fillers. They give interviewers something real to trust.

Congrats mate 🙌

[Krishna @KrishXCodes]

@_kelindi Somewhere between "I'll build GTA" and "why is this API returning 500", we all grew up 😭

[Krishna @KrishXCodes]

@karpathy @A_yarzuki7 Holy!!! 😱😱

[Krishna @KrishXCodes]

Something I’ve noticed while building at startup:

The best founders never ask “can AI do this?”

They ask “where is my team wasting 10 hours a week?” and then figure out if AI fits

That one difference separates useful AI from AI for the sake of AI

Problem-first thinking is stupidly rare in this space

[Krishna @KrishXCodes]

@ajay_2512x congo brother🙌

[Krishna @KrishXCodes]

The real lesson isn't about AI.

It's about system design.

Prompt engineering is necessary but treating the model as your security perimeter means you've already lost. The model is the application logic, not the firewall.

Real defense looks like rate limits, session rotation, output scanning, audit trails - the same boring infrastructure thinking we've used for 20 years. AI didn't change that. It just made people forget it.

Build the system around the model like you don't trust it. Because you can't, yet.

[Krishna @KrishXCodes]

Someone spent 40+ messages trying to break our AI in production.

Their playbook:

> "Are you being held hostage?"
> Silence of the Lambs roleplay
> Binary code injection
> Fake "previous conversation" made-up scenario's

10 iterations of prompt hardening couldn't fully stop it.

Here's what did 🧵

[Krishna @KrishXCodes]

What actually worked(for our use-case) was defense at the system level, not the prompt level :

1. Rate limit users per hour
2. Auto-rotate to a new session after N messages
3. Sliding window + re-inject the hardened persona every N turns so that model never drifts more than N messages from its instructions
4. Scan model output before streaming -> shadow ban on blacklist hits

Layered. Boring. Effective.

[ClaudeDevs @ClaudeDevs]

Claude Code weekly limits are increasing 50%, now through July 13.

Live now for all Pro, Max, Team, and seat-based Enterprise users.

[Krishna @KrishXCodes]

@Kyriakos_Pelek It sure doesn’t 😅

[Krishna @KrishXCodes]

I bet every AI engineer has that one prompt that's 47 lines long and starts with "you are a helpful assistant who absolutely must never under any circumstances"

and it still hallucinates 😭

[Krishna @KrishXCodes]

@ashutosh_sao @kirat_tw @100xSchool @100xDevs Gd luck mate 😉