Posted random stuff on LinkedIn, and it went like this! - 113k+ impressions within 24 hrs - 500+ followers increased - Got reached out for a gig I think I should spend more time on LinkedIn

Claude Code just got a lot more powerful. AITMPL is an open-source collection of Claude Code components. Skills, agents, commands, hooks, MCPs, plugins, loops, settings, and more. All organized by category. If you use Claude Code, this repo just became essential.

cc-lens now tells you where you're burning money on Claude Code. - Low cache hits. - Premium models on tiny sessions. - Compaction thrash. - Each with a conservative monthly $ estimate. Set a budget, get a pacing projection, see spend spikes before the bill.

Just dropped a new video 🎥 Static UIs aren't going away, but AI-native apps need interfaces that can adapt in real time. In this video, I break down Generative UI, AGUI with @CopilotKit , and why the front-end is becoming the runtime for AI. Watch here:

You can replay any Claude Code session turn by turn in cc-lens. - Tokens per turn, - tool calls, - compaction events, all there.

NVIDIA dropped a 0.6B parameter speech recognition model. Nemotron-3.5-ASR. Runs on CPU. No GPU needed. • 40+ languages • Real-time streaming • 2.5x faster than official Nemo runtime • Same accuracy Works offline. Plugs into agent pipelines. Local speech processing just got lighter and faster.

Just checked my @claudeai Code Usage in CC-lens 1 billion tokens used. Didn't even realize I hit that. Been using it daily for months now. Still the most fun I've had coding in years.

Just tried @NVIDIAAI Nemotron-3-Ultra on @nebiustf playground. Gave it a reasoning prompt. Here's what it did: • 807ms time to first token • 6.5s total time • 126.3 tokens/s For a 550B MoE model, this is fast. Worth testing if you're building agents. Demo: x.com/NVIDIAAI/st...

One thing I really like about @NousResearch Agent Desktop: You can see all the artifacts directly in the interface. No switching tabs. No downloading files. Everything's just there. Tried it on Raah.dev - worked exactly as expected. x.com/Arindam_172...

Hot take: Most developers don't have a performance problem. They have a visibility problem. They don't know WHICH users are slow, on WHICH connection, hitting WHICH endpoints. That's the actual gap. And it costs you churn you can't explain.

CC-lens Crossed 500 Stars! Thanks, everyone, for supporting this project from Day 1

Last few months taught me something unexpected. How people react as you grow. Didn't see it coming from certain people, honestly. But I'm glad it happened at a small scale. Helped me understand how things actually work.

Meetings generate a lot of “we should do this” moments 😭 So we built Brieflane that turns meeting transcripts into: - Linear issues - risks - decisions - Slack updates Built with @nextjs + @Corsair + OpenAI Agents. 100% Open Source: github.com/Studio1HQ/b... Demo below 👇

The @aiDotEngineer Singapore was incredible 🇸🇬 First time in Singapore, and it exceeded expectations. Here's my recap:

Launching something in the observability space in a couple weeks. Looking for early beta testers. If you're interested in trying it out early, drop a comment or DM. Happy to share access.

Just Shipped cc-lens v0.3.3! Key improvements: • Instant boot (~0ms) via prebuilt Next.js standalone • 7x smaller footprint • Fixed Opus 4.x pricing display ($5/$25, not $15/$75) If you're seeing lower numbers in your dashboard, that's the corrected pricing. github.com/Arindam200/...

I’m starting to feel like AI agents are exposing a flaw that already existed in engineering teams: We review changes in isolation, but systems fail through interactions. Agents just scale that mismatch faster.

I'll be attending AI Engineer Singapore! If you're there too, let's catch up!

One thing I’ve noticed while working with coding agents: The issue usually isn’t “bad code.” It’s: > retries affecting another service > timeouts shifting load patterns > one “safe” local optimization causing global instability

Just asked Codex to add some features. After 10 mins:

We've been Building a new feature. Spent days overthinking it. - Workflows, - agents, - orchestration. Then we realized, just use tool calls! No agents needed. No extra layers. Just function calls. Works perfectly.

My @cursor_ai Agents SDK video is already ranking on Google. Working on a deeper breakdown now. Stay tuned.

Awesome AI Apps is Trending on GitHub! Thanks Everyone!

Shipping AI isn’t just writing prompts. You’re dealing with retries, latency, evals, cost, observability, and unpredictable failures. It’s still software engineering. Just less deterministic.

X is wild. One week everyone's hyping Claude Code. Next week they're all on Codex. I test them myself and stick with what actually works for my workflow. You should too. Don't follow the hype cycle.

Built a real-time voice agent that listens and responds instantly. - @SarvamAI STT for speech input - GPT-OSS via @nebiustf for processing - Sarvam TTS for voice output Wired everything together with @pipecat_ai into one streaming pipeline. Runs locally. No complex setup. Full tutorial:

MCP is useful. It makes integrations cleaner. But it doesn’t fix bad system design. You still need proper permissions, safe write paths, logging, and control over what the agent can do. It’s plumbing, not a safety layer.

YC is becoming the next JEE here!

The hard problem isn't storage. It's filtering. Engram by @weaviate_io solves it at ingestion: • Topics define what gets extracted • Scopes define where it lives Memory is structured before it hits the vector DB. No dumping raw logs and hoping retrieval fixes it later.

Quoting this because more people need to see how simple this gets. From blank UI → real-time collaborative editor with comments, presence, and live sync… without touching WebSockets or CRDTs. Here’s the demo 👇 x.com/Arindam_172...

Agents don’t become reliable because the model has improved. They become reliable when the system around them improves.

100% agree. We need very niche & specific benchmarks for each task/domain. That will help us to judge what works best for this use case. But my only concern is the fact that many teams will game the benchmarks. x.com/OfficialLog...

Had a great time speaking at @gdgkolkata's Build with AI event yesterday. Built a voice agent with web access using @livekit, @GeminiApp & @olostep. The energy in the room was amazing. People were genuinely excited about the possibilities. Here's the demo I showed:

I'll be speaking at Build with AI Kolkata by @gdgkolkata tomorrow. Will walk through building real-time voice agents with Gemini 3.1 Flash + Livekit that process audio directly for natural flow 📅 Apr 25 | 9:30 AM IST 📍 SNU, A Block Seminar Hall See you there!

New update for CC-lens 👀 Just shipped: • cmd + K navigation • cleaner session view • faster access to chats & tool calls UI feels way better now. Still 100% free & open source. try: npx cc-lens

I launched Formaly as a product for creating and running better surveys. Now I am thinking more about the next layer: what happens when this becomes part of a team workflow, not just a solo workflow? That shift changes the roadmap a lot.

When people say “multi-agent systems”, most examples are just prompt chains. A real multi-agent system needs: • orchestration • service boundaries • execution tracking • governance Otherwise, it becomes impossible to debug. Projects like @AgentField_ai are trying to formalize this stack.

Improved light mode for cc-lens. Send me feedback if you have any. Will keep improving it. It's 100% free & open source. Try: npx cc-lens

Just hit 2k subscribers on YouTube 🎉 Not a huge milestone, but it means a lot. Thanks for supporting the content. Gonna be way more consistent now.

Learn How to Build Real-Time Voice Agent With @GeminiApp 3.1 Flash Live & @livekit Watch it here: www.youtube.com/watch?v=edk... x.com/Arindam_172...

Function matters. Presentation matters too. I am improving how surveys look inside Formaly.io because better design changes how seriously people take the experience. Trust starts at the interface.

I keep thinking about this: feedback should happen where the user already is. Not every survey should live behind a separate link. I am exploring more in-product feedback workflows for Formaly because that feels like a much bigger opportunity.

Infra used to be built for developers. Now it’s slowly being rebuilt for agents. Agents are now writing and running software. They need - long-running compute, - sandboxes, and - fast model access. Agentic infrastructure will unlock a new class of AI-native products x.com/rauchg/stat...

Updated my GitHub README after a while. Thoughts? (shoutout @steipete for the inspo)

We instrument everything in production. Logs, traces, metrics, dashboards, profilers. Then we pay $50-200/month for an AI coding tool that generates detailed session logs... And never look at them. Now you can, Just Run npx cc-lens

Crossed 2k Weekly downloads on CC-lens! Great to see people actually using it! Try now > npx cc-lens

Everyone in my TL got into: → VibeCon by @emergentlabs → @ycombinator Startup School India And I didn't make either. Anyone willing to help me get into one of these? Would mean a lot.

Hit my rate limits twice on a single project. @bcherny @claudeai When's the excessive token usage getting fixed? I love Claude Code. Want to keep using it. But this makes it nearly impossible to ship. ( checked the usage graph on cc-lens )

Claude Code burns tokens like a black box. Built CC-lens so you can see where they actually go. Project spend, cache hits, tool usage, and full session replay. Recorded a full walkthrough if you want the details. x.com/Arindam_172...

Alibaba's Qwen 3.6-Plus just matched Claude Opus on coding. Half the size. 1M context. Built for agents. It's beating Kimi K2.5 and Gemini 3 Pro on SWE-Bench. Free on OpenRouter right now. Open source soon. Smaller models are finally catching up. x.com/Alibaba_Qwe...