Gaurav Kumar

Excellent Opportunity to learn how @JioHotstar scales up 📈 in this 1st edition of ScaleUP, I will be presenting how we do Data@WC scale in 8-bit style 🕹️

@AsfiShaheen Thanks 🙌

I get it. Here’s what I did with tool calls 1. After writing my gold set I treat tool definitions similar to schema ie fact. So there’s a tool in one of my signatures for exploring nearest neighbor of a target page. The description I take as fact 2. What gets optimized is the second signature which has access to this tool. My gold set has already specified situations on when to call this tool. All GEPA now does is morph the instruction text of the signature which decides when to call this tool that’s why I find the fact vs instruction difference helpful. Worth mentioning tool definitions are a sort of hybrid. I have actually played with running GEPA on tool definitions themselves but wasn’t as useful.

GEPA learnings: optimize instructions, freeze facts A few months ago I bumped into GEPA. Was overjoyed because it helped me tag close to a million PDF pages accurately with a low cost LLM. So I pushed data set on psxGPT, a version worked, got to a 1,000 users. But then my serious users started asking for structured data and ability to model more easily. I added those, hand crafted some rules to handle routing complexity, worked for a while but reliability sucked so I reverted to older commit. As I started debugging and re creating my backend in a single Jupyter notebook and re running evals, gaps became clearer. Signature boundaries weren’t tightly defined. As I closed those one question remained: What should GEPA optimize and what should it leave alone? Where i ended: freeze facts, optimize all else. Previously I never touched Signature docstring. But I notice letting GEPA change docstring is super effective (DSPy Signatures have a little docstring instruction in addition to input and output). For me what vs how distinction isn’t helpful. Fact vs instruction is clearer. So schema is fact, column name fact, ticker set fact, list of canonical fact. Everything else: instruction and therefore should be optimized. This has also led to me designing my context folder by Signature and separated by schema.md (which I don’t optimize) and a bunch of others which I do. This process has raised another important question When is the right time to run GEPA? Only when signature boundaries are super tight and you’re very sure no obvious instruction is left out. Meaning whatever error is arising is just model drift and not a function of your own instruction sloppiness. So for instance: if a stage requires inserting canonical names of a bank’s balance sheet and you didn’t insert those only to see model getting lost, fix that first. Writing the spec, tests, defining signature boundaries, manually checking code blocks in Jupyter notebook is tedious yes but essential. I like to rebuild until cognitive load vanishes. That’s when I know that I actually “know”. Otherwise I’m winging it and that feels unsettling. GEPA now has an API which is very easy to use. I prefer it over using the DSPy library. There are some similar optimizers out there too. I don’t understand the underlying genetic Pareto stuff well enough but this API works very well for me. Swapping DeepSeek V3.2 for Gemini 3.1 flash lite was a breeze.

@AsfiShaheen I have good experience with DSPy and some with GEPA but felt that the current way of improving tool call usages/descriptions is bit hand wavy.

No blog or code to share sir, it’s deeply embedded in a private repo. In any case all you need for every signature (including tool calls) is ~25 gold set examples of what correct looks like. After that it’s plug and play. Give your coding program their api docs and your gold set examples, rest will be very straight forward.

So many people give over-emphasis to knowledge. It can easily be attained, especially in today's world. In my experience, I always prefer Agency >> Skill >> Stack familiarity If the candidate has high agency, they will be able to acquire any skill and learn any tech stack.

@lateinteraction @tech_optimist Simplest way is to think of it like ProgramOfThought module but applied to prompts as well

@tech_optimist RLMs can print(prompt) or copy prompt[11:17] or prompt.count(“r”) because their prompt is a variable. An object. A symbolic reference. It solves this issue: x.com/davidsholz/sta…

Would RLMs struggle with this too?

@getpy @DSPyOSS @langchain @crewAIInc @pydantic Survivorship Bias 😊

Which AI framework do you use?. If your option is Others, which one?. And Why?. Please RT @DSPyOSS @langchain @crewAIInc @pydantic

Hyrum's Law: with a sufficient number of users of an API, it does not matter what you promise in the contract: all observable behaviors of your system will be depended on by somebody. Making Postgres connections faster broke implicit rate limiters!

@mitsuhiko Try DSPy

I built stuff now with Claude code sdk, vercel AI SDK and pedantic AI and I found them all quite good, but annoying in their own rights when trying to work with serialized context. Vercel AI SDK might be the best on I tried when driving the loop manually.

@neelesh_salian All they need is a spark!

Confluent CFLT Stock is way down. Flink might not be working out.

@dhh @tahayvr Theme of the month should get $100 gift card by @dhh

Not a day has gone by recently where I haven't woken up to a cool new Omarchy theme created by someone in the community. This Gold Rush look is just right for the morning sun. Thanks @tahayvr! github.com/tahayvr/omarch…

We at @tursodatabase have just done something unheard of in the startup world: we hired a person in prison. Not an ex-con: he is actually serving time right now. If you follow us you may remember when we announced our full rewrite of SQLite in Rust, Project Limbo. It got a tremendous amount of attention, but by far, my favorite story was that of @PThorpe92: Preston is currently incarcerated for things he regret doing a decade ago. A model reformed inmate, Preston was given the beautiful opportunity by the Maine Department of Corrections to access the internet. He then found our project, and since it's not like there are many entertainment options competing for his attention in prison, he decided to pour his heart and soul into helping us rewrite SQLite. I immediately fell in love with his story. Reading his old blogs, his Github profile. I prayed we would be able to hire him as soon as he was out, and make him a part of our team. But thankfully, things have happened in such way, that we were able to do it *even before* he was released. He's now given an opportunity to make a dent in the world, from prison. Preston claims he is immensely thankful for this opportunity. It is certainly not an opportunity that comes often, and I understand him. But when I read his story in his personal blog, where he mentioned how he clearly saw his life going downhill, had this sudden epiphany, saw clearly that he could do differently, and then started being showered with blessings, I saw the unmistakable markings of how our Lord Jesus Christ, the Lamb of God, operates in the world. I was an obnoxious Richard Dawkins-style atheist for more than two decades. I am not an experiential person, and I came to a change through the rational evaluation of the historical context of His resurrection. I am not the kind of person that feels the hand of God in my life, but over the years I learned to recognize his patchwork through those sudden, unexpected changes, where everything suddenly "just works". And because of that, I am the one who's thankful. The Lord decided to operate such a drastic change in Preston's life. Clearly He has a great plan for him. I am immensely thankful for the opportunity to play a small part in that story. God, in his infinite power, doesn't need me to do any of that, and yet he gave me the opportunity to act. I feel thankful and blessed. I am also very thankful to @LabsUnlocked , and entity helping people like Preston find a better life once they're out. If you feel inspired by their mission, you should reach out and help. And if you want to hear from Preston in his own words, read the article he just published on the Turso blog: turso.tech/blog/working-o…

@OnlyXuanwo x.com/cereal_learner…

Talk is cheap. Show me the prompt.

@d4m1n Deepseek had a lot of thinking in Chinese actually. When doing deep thinking, it would turn to think in Chinese and then finally comes out to give the conclusion in English again. Likely because it was trained on such a training set.

galaxy brain question 🧠 is using ChatGPT in Chinese cheaper because it's using fewer tokens?

@jasonfried Let's move this to backlog and we'll prioritize it later ... and other lies we tell ourselves

There's no such thing as low priority. It's either high priority, or no priority ("it just is"). Multi-level priority systems instituted in software is software doing what it does worst.

Can someone recommend a good UUID? I've searched a lot and a8098c1a-f86e-11da-bd1a-00112444be1e seems pretty random. Want to make sure no one else has used it before!

@shuttle_dev End to End it takes a lot of time. I mean, deploying is easy but first we have to rewrite it in Rust

What’s the biggest pain point you’ve faced trying to deploy a backend application?

@github WHY?

@abhicantdraw Underrated

Meanwhile (many) CBSE students choose Redis because they owe their education to KVs.