Francesco Pochetti

AWS ML Hero. MLE @boltapp. Failed Chemist. Blogging about my ML/DL journey. IceVision core-dev.

[Francesco Pochetti @Fra_Pochetti]

If you have trained a model, you want to deploy it to SageMaker, and your first instinct is to start screaming, please😌 and keep reading.
Today, I am launching a course: How to deploy an ML model to SageMaker with a custom Docker container
👉 Get it here: https://t.co/ENPVDRwgkx

[Francesco Pochetti @Fra_Pochetti]

@thealokverse @Yuchenj_UW Got demoted by Gemini

[Francesco Pochetti @Fra_Pochetti]

The full story covers: disasters, pivots, technical architecture, and what actually works when vibe-coding with LLMs.

Tech: Next.js 15 • React 19 • Prisma • Replicate • AWS Bedrock

#AI #LLMs #VibeCoding #BuildInPublic

[Francesco Pochetti @Fra_Pochetti]

I finally shipped Visual Neurons: a personal AI studio for image/video generation.

9 models. One interface. Zero lines of code written by me.

Claude Sonnet 4.5 + Cursor built it.

📖 Full story: https://t.co/4qBGIuaP6U
🎥 Demo: https://t.co/WUwYPEvhhm

Here's what it took 🧵

[Francesco Pochetti @Fra_Pochetti]

3 complete pivots:

Feb: AWS Amplify (15-min builds, mysterious failures)
Apr: Vercel + Supabase (too many platforms, blind LLM debugging)
Aug: EC2 + local dev → shipped

The breakthrough: KISS. No S3, no complexity. Just Next.js + SQLite + nginx.

[Francesco Pochetti @Fra_Pochetti]

Working on adding Runway Gen-4 video generation.
Close but not there yet.
https://t.co/TiPkUgIL44

[Francesco Pochetti @Fra_Pochetti]

🎉🎉 Seems I got Gemini Image Edits to work! 🎉 🎉

[Francesco Pochetti @Fra_Pochetti]

Getting close to Chat with Gemini integration

[Francesco Pochetti @Fra_Pochetti]

@richdevelops Yes!

[Francesco Pochetti @Fra_Pochetti]

When I thought I was done with outpainting, I started stumbling on frontend timeouts. The StabilityAI API call is successful though (I see it in 💸). Lambda logs confirm too.

Seems AppSync plays tricks on me. Here's how ChatGPT puts it 👇 (and yes, I have to move to async)

[Francesco Pochetti @Fra_Pochetti]

Stability's outpainting is not bad at all.
300px completely "invented" in all directions.

[Francesco Pochetti @Fra_Pochetti]

What happens in that scenario?

Stability finishes the job in, say, 40 seconds, returns HTTP 200, and you get billed.

However, AppSync has already cut off your resolver at 30 seconds, so your front end sees a “timeout” or “network error,” rather than the final 200 result.
```

[Francesco Pochetti @Fra_Pochetti]

```
If you are calling the Stability outpainting API from an AWS Lambda resolver (via AppSync), remember that:
AppSync imposes a 30-s timeout for all resolver calls. Even though your Lambda function can be configured for a 120-s timeout, AppSync itself will cut off at 30 seconds.

[Francesco Pochetti @Fra_Pochetti]

Managed to add Outpainting from StabilityAI 🎉
Now hardcoding 100px in each direction.
https://t.co/5fSXBo1w3z

[Francesco Pochetti @Fra_Pochetti]

Seems to work

[Francesco Pochetti @Fra_Pochetti]

I am close for Style Transfer from Stability.
It'll happen tomorrow.

[Francesco Pochetti @Fra_Pochetti]

I am close for Style Transfer from Stability.
It'll happen tomorrow.

[Francesco Pochetti @Fra_Pochetti]

Integrated StabilityAI image generation.
After the 20-commit-3-days refactor it was a breeze 🙌
It boiled down to implementing the `generateImage` func inside the `stabilityProvider.ts` and the dispatcher propagates to Lambda and the frontend.

https://t.co/FNHJkDVogE

[Francesco Pochetti @Fra_Pochetti]

Claude is happy with the refactoring 🫡

https://t.co/RR6dvqblqJ

[Francesco Pochetti @Fra_Pochetti]

Weekend plans!
Believe it or not but it's not fully refactored yet 😱

[Francesco Pochetti @Fra_Pochetti]

@robmarkcole @AnthropicAI `o3-mini-high` absolutely kills it. Claude 3.7 is very very good but it requires too much fiddling with the context. Super annoying.
Gemini image understanding is top notch (together with GPT4 tbh). Never really given it a proper shot as a coding assistant

[Francesco Pochetti @Fra_Pochetti]

Not sure what they are doing at @AnthropicAI, but Claude is basically unusable for any reasonably long context.
ChatGPT handles the same queries without any issues.