News for hackers. Stuff that matters.

manidoraisamy

The ones that are gone are comments/documentation in code, unit testing, design patterns, datamodel/db design. I am happy documentation is gone. But I miss design, especially db design. AI generates SQL/code with no regard for backward compatibility or future maintenance.
Not as much as US. Everyone from trump to Altman is betting on fusion. But China is more pragmatic by focusing on making fission resistant to supply chain shocks in uranium. Since they are fast follower, their plan might be to catch up, once fusion is viable for practical use.
Re:Show HN: Send email to fill Google Forms on Wednesday July 02, 2025 @12:42PM (#44445843)
Yes. We use Gemini with mcp which supports photos and PDFs natively.
Ideally, you want to start small and iterate. With Promptrepo, you can use versioning to compare model outputs across different datasets. In the test UI, we calculate confidence scores using @promptrepo/score [1], which parses OpenAI’s logprobs and shows field-level reliability. Fields with low confidence are highlighted in red, making it easy to catch signs of overfitting or data drift.

[1] https://github.com/ManiDoraisamy/promptrepo-score

Our target customers are product teams like ours. I’m the technical founder and was building a feature to extract product prices from unstructured text. But my co-founder, who understands all the pricing formats — is non-technical. I built Promptrepo, so he could finetune the model himself and improve accuracy without relying on me. My bet is that this same dynamic exists in many other product teams.

Thanks again for the feedback and recommendation. We’ll update our site with clearer examples and target use cases!

Finetuning is technical, but OpenAI hides things like layers, learning rates, and uses LoRA under the hood. You just upload examples - usually around 50.

But even that’s too much for most business users. Choosing input/output fields, generating and validating JSONL still feels like coding. That’s why we built PromptRepo: it turns finetuning into a guided, no-code workflow using spreadsheets.

Thanks for the clarification. OpenAI projects in ChatGPT are meant for end users to get personalized help using their own documents, inside the ChatGPT UI.

Promptrepo is for developers and product teams to build new AI-powered features in their product. It’s about creating custom models that run behind the scenes in apps, not just improving a personal chat experience.

So while OpenAI projects use RAG for better chats, Promptrepo helps teams build and deploy fine-tuned APIs that serve structured outputs like JSON, labels, or extracted fields to build your own AI powered product.

It depends on the use case. For many business workflows, where structured data is key - spreadsheets are already the source of truth. But for chat-based or unstructured tasks, a custom UI might make more sense.
Can you please clarify what you mean by "OpenAI projects"? Are you referring to the playground or the API for prompting or fine-tuning?
Yeah, just like Dropbox was a passthrough for aws s3.

Edit: Sorry about the snide comment. But if this ends up as a simple utility for finetuning, I would be happy with that too. Just want to share a tool that's been very useful in building ai features in our products.

We don’t have automated evals, latency, or cost comparisons yet. But, Promptrepo does offer versioning and lets you deploy the same model across providers for comparison. Automating these comparisons is definitely on our roadmap.
Exactly! Finetuning needs at least 10 examples to even work. That’s why Promptrepo begins with prompting and schema-based generation when teams have little or no data. As they gather more examples, it gradually shifts to fine-tuning. It’s the classic cold start problem and we’ve simplified it for product teams who want to launch quickly but improve accuracy over time.

Older »