Ravi

AI/ML Engineering @ClickhouseDB
prev. @ZalandoTech

[Ravi @ravikml]

Tinkering with agents having a tool to switch complexity levels.

The same tail appended LLM context sent to different size of LLMs seems much more tractable than having arbitrary rules like “plan with fable and write an MD file and hand it to GPT 5.5 for implementation”

[Aaron Katz @ceo_clickhouse]

We built a custom sales co-pilot on ClickHouse Agents, and 81% of our sales org uses it every single day to manage accounts worth hundreds of millions of dollars. They're having thousands of conversations that use our data warehouse, salesforce, internal docs and public feeds to support our customers.

We have observability of every conversation via Langfuse, so we can grade answer quality, optimise model selection and pass intelligence to other teams on where sales needs support.

It's built entirely on the ClickHouse Cloud stack using @ClickHouseDB , @LibreChatAI and @langfuse .

[Aaron Katz @ceo_clickhouse]

For the first time, an open weight model (GLM5.2) made it into the top 5 of our benchmarks for agentic analytics, and is well within the threshold to be seriously usable. I'm excited to see where that takes us.

[David Cramer @zeeg]

GitHub friends: it'd be great to have a way, via the API/CLI, to upload photos to issues/pull requests.

AFAICT the only way to do it right now is browser emulating or hosting the content somewhere outside of GitHub, which means having agents help QA/upload visual artifacts sucks

[Ravi @ravikml]

@brian_armstrong Very cool stuff! Glad to know you guys use librechat. I work on it’s features and maintenance @ClickHouseDB

Let me know if we can help with making it more cost effective for you

[Ravi @ravikml]

@BenceRedmond I ask codex to run claude -p “command to diff +summary” or cursor’s agent for faster reviews.

Cursor’s skills for code review in their team kit are also pretty great! I use it with every agent!

[Ravi @ravikml]

@benhylak It does a similar job for much less output tokens? That’s still a brag

[Ravi @ravikml]

Launching codiff through the codex/claude before commiting is an absolute find!!

https://t.co/hTTjNSbTFm

[Ravi @ravikml]

@theo End to end validation with repeatable custom skills that run on docker compose or kind local k8s

[Ravi @ravikml]

@theo Cursor’s thermo-nuclear-code-quality-review to find maintainability issues.

Claude review to find lateral thinking ideas to simplify and throw away components while keeping the functionality.

[Ravi @ravikml]

@gokulr Neural diversity ftw!

[Ravi @ravikml]

Finding arm64 images to run on a mac is a pain. I might set this up on a remote box on hetzner or something.

[Ravi @ravikml]

https://t.co/mW60Gt2N11

Local k8s on docker. This is a massive unlock. Validating k8s setups I have using codex/claude locally is 😚🤌

[Ravi @ravikml]

This 💯

Can’t wait to write workflows in skills and have them actually run reliably

[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.

[Claude @claudeai]

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.

Available today at the same price.

[Ravi @ravikml]

I had a decent grasp of how LLM inference works. But I hadn't connected a lot of dots on how this can work in production where cost, throughput and latency matter. So I did a bunch of reading and wrote down what I understand.

https://t.co/CnzKnBWTqi

[Ravi @ravikml]

A lot of programming wisdom like “keep things extensible” don’t seem to apply anymore when you can sling thousands of lines of code every minute.

“Keep things verifiable” - write tests, precise types, benchmarks etc. still stands though

[Gary Bernhardt @garybernhardt]

I see that "types and tests do the same thing" has once again reared its head, so here's Ideology again: https://t.co/4iCRhd2Vfq

[Ravi @ravikml]

@dhh At the limit, you’ll end up writing what looks like a type checker in your test suite.

[Ravi @ravikml]

Starting to use codex desktop way more than Claude. Computer use through the ChatGPT mobile app is a game changer!