For the last 30 days I’ve gone deep on the infrastructure behind AI agents. Not surface-level. The actual internals. Here’s everything I studied, so I can contribute to the companies building this stack: MCP (@AnthropicAI ecosystem) → JSON-RPC 2.0 message framing, request/response lifecycle → the tool / resource / prompt model and how a server declares capabilities → stdio vs HTTP/SSE transports → schema design with Zod (TS) and Pydantic (Python) → the TypeScript and Python SDKs, plus the Go and Rust SDKs → the SEP proposal process for spec changes Agent frameworks (@LangChainAI, Microsoft AutoGen) → the Runnable / LCEL abstraction and how it composes → async Python patterns and Pydantic v2 internals → how integrations (LLMs, vector stores, tools) are structured → LangGraph’s graph/state model → multi-agent orchestration and message routing between agents Inference engines (@vllm_project, @ollama) → paged attention and KV-cache management → continuous batching and request scheduling → quantization → Ollama’s model runner architecture and the GGUF format → the CGo bridge to llama.cpp ML libraries (@huggingface transformers, PyTorch) → how a model is structured: config, modeling_*, tokenizer → the “add a model” and “fix a model” patterns → autograd and the training loop → PyTorch’s ATen core and where Python ends and C++/CUDA begins → tokenizers, datasets, accelerate Evals (@OpenAI evals) → how an eval is specified: YAML registry + Python logic → prompt and grading design → the eval-template pattern for measuring whether a model is actually good That’s the study phase. Now the real part. Next 30 days: one open-source PR a day, to the projects I just listed. MCP, @huggingface, @LangChainAI, @vllm_project, @ollama, @OpenAI - I’m coming for your repos. See you in September.

I’m raising funds to make reconciliation free for tax filing in the US, starting with Schedule 5. Because I don’t want to build another reconciliation business and make money off something that should just work. If you’re smart enough to see what I’m building, DMs are open.

Next week is going to be a big one!! Two product launches coming up. 👀

Staff accountants will get fucked

Hi @isro , I have few satellites to send

Joined the Kimia waitlist #0065 Your position deserves protection. Get your custom @KimiaProtocol pack. Claim your spot today 👇 waitlist.kimia.live/?pack=ki…

Go go go

Most people enter the market unprotected Join the Kimia waitlist and get your own digital condom waitlist.kimia.live

Tag the most influence people you know in bangalore btw is this India?

AI can’t build you taste

Btw, I’m cooking something, Was out of my OSS contribution because of this, but I’ll ship the demo video tomorrow!! Also, used Jev, it’s pretty cool. Good work, @typesafeai team!

Arakis Kickoff #2 has entered the chat. 🟠⚽️ New faces. New energy. Same crazy mission. We’ve got a lot to build, a lot to break, and probably way too many football arguments ahead 😂 Let’s see what happens. 👀

Traded my way to the top Took the W in the @arakisdotfun trading competition. From on-chain PnL to holding the prize IRL. GG @arakisdotfun 🫡 If you haven’t checked out what they’re building yet, go play around with Arakis: on-chain prediction markets on Solana. arakis.fun cc: @HeySunit @imanishbarnwal

Can’t agree more

Now I think KBC is the easiest way to raise pre-seed funding What traction? What PFM? Just answer 10 question

My broody is looking for backend/full-stack opportunities in Bangalore. Please share or reach out if you’re hiring! 🙏

Just got off a call with @lightspeedvp

This is the kind of AI progress that actually matters. Not just bigger models fundamentally better economics: faster inference, dramatically lower cost, and intelligence built for making decisions. If this scales, the bottleneck shifts from “can AI do it?” to “what should we let AI do?”

Day 8. New week, new repo. After evals turned out to be unmaintained, I moved to openai-agents-python, the one that actually merges outside contributors. Found and fixed a data-loss bug in the shell tool 🧵 It was truncating output twice against the same budget, so the command label itself ate into max_output_length and silently chopped the real stdout. With a small budget, the model saw only the prompt prefix, none of the actual output. Fix + regression test: github.com/openai/openai-age… @OpenAIDevs