by veilgen on 3/18/2025, 7:56:48 AM
This looks like a powerful step toward making prompt engineering more scalable and production-ready. The version control approach, along with staging environments and real-time analytics, seems particularly useful for teams handling high-volume AI workloads.
One question: How do you handle prompt drift over time? As models evolve, prompt effectiveness can degrade—do you provide any automated testing or monitoring to detect when a deployed prompt needs adjustment?
Looking forward to exploring Portkey’s capabilities.
by JTyQZSnP3cQGa8B on 3/18/2025, 8:04:18 AM
> How do you version control thousands of prompts?
Kill me now.
> How do you collaborate across hundreds of engineers
What do you mean by that? This only targets a few big companies.
> A tech firm that cut deployment times from 3 days to near-instant
That's a process and maybe CI issue, I don't see how AI would improve any of that but I'll be gladly proven wrong.
> You can try it yourself at prompt.new
All I see is a login page from another company. Don't you have a web site with all those serious prompting you do?
We've all seen the explosion of prompt engineering tools lately. While playing around in playgrounds is fun, when we tried to take our AI prompts to production, we hit a wall. I'm guessing many of you have experienced similar pain points.
We kept hitting questions nobody seemed to be answering: How do you version control thousands of prompts? How do you handle multiple production deployments? How do you scale from prototype to millions of requests per day? How do you collaborate across hundreds of engineers without stepping on each other's toes?
So we built Portkey's Prompt Engineering Studio - a complete toolkit designed specifically for productionizing AI prompts across 1600+ models.
Some technical details that make our approach different:
- High-performance infrastructure: We've deployed prompts as large as 500,000 tokens with production-level latency - Git-like version control with instant rollbacks for prompt deployments - Mustache templating system for parameterization and reusable snippets - Publish/release flow with proper dev/staging/prod environments - Real-time analytics tracking prompt performance, latency, and token usage - Native integrations with Langchain, Llamaindex, and Promptfoo
The scaling capabilities have enabled some impressive use cases:
- A content company running 500+ prompts across 700+ websites - A tech firm that cut deployment times from 3 days to near-instant - Education platforms with hundreds of non-technical creators building AI workflows
Our platform has processed hundreds of millions of prompt completion requests already, with over 10,000 prompts deployed to production environments.
We think the HN community will especially appreciate our approach to bringing software engineering best practices to AI development!
You can try it yourself at prompt.new
I'd genuinely love to hear how others in the community are handling these challenges, what you think of our approach, or any other feedback you might have. This community has been invaluable in shaping how we think about developer tools.