Today’s lead · x.com
vLLM 0.25.0 makes Model Runner V2 the default and retires legacy PagedAttention
vLLM released v0.25.0 with 558 commits from 232 contributors. Highlights include Model Runner V2 as the default for dense models, retirement of the legacy PagedAttention implementation, a native-speed Transformers backend, a unified Streaming Parser Engine, universal speculative decoding across heterogeneous vocabularies, DSpark and DFlash drafters, and Hy3 and Unlimited OCR model support.
Top signals
1 moreTools & repos
5 selectedDicklesworthstone/destructive_command_guard
A Rust guard for blocking dangerous git and shell commands before AI agents execute them. It is a practical safety layer for teams letting coding agents run terminal commands in real repositories.
Graphify
Graphify maps code, docs, schemas and other project artifacts into a local queryable knowledge graph so coding agents can ask about relationships instead of repeatedly grepping files. The input describes tree-sitter parsing for code, EXTRACTED/INFERRED/AMBIGUOUS edge labels, MCP exposure, and hooks for assistants such as Claude Code, Codex and Cursor.
FetchSandbox
FetchSandbox is an API integration testing tool focused on retaining failure history. It is useful when teams or agents are wiring third-party APIs and need regressions surfaced instead of rediscovered by hand.
Codex plugin for Claude Code
The plugin lets developers use Codex from inside Claude Code for code reviews or delegated coding tasks, according to the input post. It points to a practical direction for multi-agent coding workflows: less copy-paste handoff, more IDE and CLI plumbing.
Second Brain for AI v2
Second Brain for AI v2 offers persistent AI memory intended to connect context across tools rather than leaving each assistant session isolated. It is worth watching for teams repeatedly reloading project context, though the real value will depend on integrations and permission controls.
Blogs worth your time
4 reads
Build a containerized Python MCP server for Claude Code
Balajee Asish Brahmandam walks through building a Python MCP server, testing it locally, packaging it with Docker, and connecting it to Claude Code from the terminal. The useful bit is the production framing: reproducible deployment, security risks, and handoff-ready commands instead of a Claude Desktop screenshot demo.
Building Reliable Agentic AI Systems
Thoughtworks’ PRINCE case study explains how Bayer moved from keyword search to an agentic RAG and Text-to-SQL platform for preclinical safety data. It is useful because it treats agents as engineered systems: bounded workflows, context routing, citations, reflection checks, observability, failure handling, evaluation and human-in-the-loop controls.
Ragged Paged Attention on TPU: design philosophy and a vLLM case study
This video and blog explain TPU memory hierarchy, tiling, packing, single-threaded double-buffering, and why ragged paged attention needs different kernels from GPU-oriented approaches. It is useful for inference engineers dealing with dynamic sequence lengths and KV-cache updates on TPUs rather than assuming GPU tricks transfer directly.
Building LLM decode kernels from scratch on an RTX 3060
The thread implements GEMV, RMSNorm, SiLU, softmax and attention kernels from naive to optimized, checking them against an fp64 CPU reference. The concrete measurements—GEMV at roughly 92% of DRAM peak while using about 1% of compute, and flash-decoding near 80% of peak at 16K context—make the memory-bound nature of decode tangible.
Funding & acquisitions
3 moves
Indian startup IPO pipeline
Inc42 says 28 Indian startups have filed DRHPs with SEBI and more than 24 are finalising IPO plans, with Zepto, OYO, InMobi and Zetwerk alone potentially raising over ₹45,000 Cr in 2026. The signal for founders is clear: public-market appetite is shifting toward predictable cash flows, sustainable unit economics, governance and low burn rather than headline growth.
Indian startup VC inflow
YourStory reports Indian startups raised $107M in the second week of July, the second straight week of muted funding activity. The largest round was Adage Automation’s $24M raise, and the absence of large-ticket deals kept weekly venture inflows subdued despite continued interest in deeptech and AI.
Aukera
Aukera raised $10M in funding co-led by Alteria Capital, less than twelve months after closing a $15M equity round. It is one of the few concrete India deal signals in an otherwise muted funding week.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.