DEV Community: Rohit Raj

Block's Buzz (2026 Guide): Self-Host the Workspace Where AI Agents Are Teammates, Not Bots

Rohit Raj — Fri, 24 Jul 2026 06:38:34 +0000

Block released Buzz on July 21, 2026 — an Apache-2.0, self-hostable workspace built on Nostr where AI agents join channels as cryptographically-signed members, not permission-restricted bots. It hit 7,600+ GitHub stars in three days. The launch coverage tells you what it is; this guide shows you how to actually run it: the exact install path, how to onboard a Claude Code or Codex agent with its own keypair, where Buzz genuinely beats Slack-plus-bots, the compliance gaps that should keep it out of production today, and the hardening checklist the same week's OpenAI–Hugging Face incident makes non-negotiable.

Read the full version with code samples, diagrams, and architecture details: Block's Buzz (2026 Guide): Self-Host the Workspace Where AI Agents Are Teammates, Not Bots

More engineering notes: rohitraj.tech/en/notes

Gemini 3.6 Flash vs 3.5 Flash-Lite: Which One to Ship — and the Price Hike Nobody Leads With (2026)

Rohit Raj — Thu, 23 Jul 2026 11:43:35 +0000

Google's July 21 drop is an efficiency release, not an intelligence release: Gemini 3.6 Flash scores the same Intelligence Index as 3.5 Flash but finishes tasks in half the time at a lower per-task cost — while Flash-Lite quietly got a 67% output-price increase. Here's the real per-task math, the migration code, and the tier decision I'd actually ship.

Read the full version with code samples, diagrams, and architecture details: Gemini 3.6 Flash vs 3.5 Flash-Lite: Which One to Ship — and the Price Hike Nobody Leads With (2026)

More engineering notes: rohitraj.tech/en/notes

This Week in AI Dev: Kimi K3 Lands and Cursor Springs a 0-Day (Week 30 of 2026)

Rohit Raj — Tue, 21 Jul 2026 06:00:46 +0000

Week 30 of 2026 split in two: China's labs shipped trillion-parameter frontier models while the tools that run agents got a hard security look. Moonshot's Kimi K3 (2.8T params, open weights by July 27) landed at #2 behind Claude Fable 5, Alibaba previewed a 2.4T Qwen 3.8, and xAI's grok-build hit 20,982 stars in under two weeks. Meanwhile an unpatched Cursor 0-day went public after seven months, Codex started encrypting sub-agent prompts, and Claude Code quietly moved to Bun-in-Rust.

Read the full version with code samples, diagrams, and architecture details: This Week in AI Dev: Kimi K3 Lands and Cursor Springs a 0-Day (Week 30 of 2026)

More engineering notes: rohitraj.tech/en/notes

OmniRoute Review (2026): Is the 20k-Star Free AI Gateway Worth It vs OpenRouter & LiteLLM?

Rohit Raj — Mon, 20 Jul 2026 14:12:06 +0000

OmniRoute is the AI gateway that shot past 20,000 GitHub stars in days: one local, MIT-licensed, OpenAI-compatible endpoint that fans out to 268 providers and 500+ models, with an 18-strategy fallback engine and 15-95% token compression. The best part is real — it runs 100% on your machine with your own keys and never phones home. But the "1.4 billion free tokens" headline, the TLS-fingerprint stealth, and the Cursor-intercepting MITM proxy are exactly the features a careful engineer should treat with suspicion. This is the honest review: what OmniRoute genuinely does well, where it beats OpenRouter, LiteLLM, and Portkey, the failure modes the promo posts skip, and precisely when I would — and would not — put it in a workflow.

Read the full version with code samples, diagrams, and architecture details: OmniRoute Review (2026): Is the 20k-Star Free AI Gateway Worth It vs OpenRouter & LiteLLM?

More engineering notes: rohitraj.tech/en/notes

MCP Goes Stateless: Migrate Your Server Before the 2026-07-28 Spec

Rohit Raj — Sun, 19 Jul 2026 04:50:31 +0000

The MCP 2026-07-28 specification goes final on July 28, 2026, and it rewrites the protocol to be stateless: no more initialize handshake (SEP-2575), no Mcp-Session-Id header (SEP-2567), with protocol version and client info moving into a _meta field on every request. That one change lets MCP servers deploy like any stateless service — serverless and Kubernetes autoscaling finally work without sticky sessions — but it breaks every server that assumed a session. Here is the full before/after migration in TypeScript: the stateless transport config, per-request _meta, the new Tasks extension lifecycle, the -32602 error change, the six auth-hardening SEPs, and exactly which servers should wait.

Read the full version with code samples, diagrams, and architecture details: MCP Goes Stateless: Migrate Your Server Before the 2026-07-28 Spec

More engineering notes: rohitraj.tech/en/notes

The Anti-AI-Slop Design Skill: How Hallmark Fixes Generic AI UI in 2026

Rohit Raj — Sat, 18 Jul 2026 09:31:33 +0000

Every site your AI coding agent builds looks the same: Inter font, a purple gradient, six identical cards, a bounce on every hover. Hallmark — a design skill for Claude Code, Cursor, and Codex that hit 12.4k stars this week — runs 57 "slop-test gates" to refuse those defaults before the code is emitted. Here is what AI slop actually is, exactly how Hallmark works, the four verbs with real commands, how it stacks up against frontend-design, Impeccable, and Stitch, and when a skill still will not save you from a bad design.

Read the full version with code samples, diagrams, and architecture details: The Anti-AI-Slop Design Skill: How Hallmark Fixes Generic AI UI in 2026

More engineering notes: rohitraj.tech/en/notes

Apple SpeechAnalyzer vs Whisper: On-Device Speech-to-Text in 2026

Rohit Raj — Fri, 17 Jul 2026 04:24:53 +0000

Apple shipped SpeechAnalyzer in iOS 26 and macOS 26 with zero published accuracy numbers. The first rigorous benchmark just landed: 2.12% word error rate on clean English, beating every on-device Whisper model and running ~3x faster than Whisper Small on an M2 Pro. Here is the full Apple vs Whisper vs Parakeet vs Qwen3 breakdown, the Swift to wire it up, the speaker-diarization gap nobody mentions, and exactly when you should still reach for Whisper.

Read the full version with code samples, diagrams, and architecture details: Apple SpeechAnalyzer vs Whisper: On-Device Speech-to-Text in 2026

More engineering notes: rohitraj.tech/en/notes

Inkling 975B: The Open-Weights Model Almost Nobody Should Self-Host (2026)

Rohit Raj — Thu, 16 Jul 2026 04:12:15 +0000

Thinking Machines released Inkling on July 15, 2026 — 975B params, 41B active, Apache 2.0, 1M context, weights on Hugging Face. Every writeup tells you how to run it. None tells you whether to. The BF16 checkpoint needs 2 TB of VRAM; NVFP4 needs 600 GB. The 8x H200 box they name is an AWS p5en.48xlarge at $63.296/hr — $46,206/month always-on. Against the $4.68/M output API, self-hosting breaks even at 9.87 billion output tokens a month. Here is the VRAM ladder, the real cost math, the July 17 price hike everyone missed, and the quant trap that will eat your agent.

Read the full version with code samples, diagrams, and architecture details: Inkling 975B: The Open-Weights Model Almost Nobody Should Self-Host (2026)

More engineering notes: rohitraj.tech/en/notes

Bonsai 27B: A 27B Model on Your Phone — and the One Benchmark That Collapses (2026)

Rohit Raj — Wed, 15 Jul 2026 05:02:29 +0000

PrismML shipped 1-bit and ternary builds of Qwen3.6-27B on July 14, 2026 — 5.9 GB for ternary, 3.9 GB for 1-bit, running at 163 tok/s on an RTX 5090 and 11 tok/s on an iPhone 17 Pro. Every writeup leads with "retains 95% of baseline." Nobody breaks out the row that matters: tool-calling drops 80.0 to 66.0 at 1-bit — degrading 4.6x worse than math. For a model sold on laptop-local agents, that is the whole story. Here is the variant decision table, the runnable commands, the KV-cache trap that makes 5.9 GB of weights need 13.7 GB of RAM, and how I would ship this in production.

Read the full version with code samples, diagrams, and architecture details: Bonsai 27B: A 27B Model on Your Phone — and the One Benchmark That Collapses (2026)

More engineering notes: rohitraj.tech/en/notes

This Week in AI Dev: The Agent Fleet Grows Up (Week 29 of 2026)

Rohit Raj — Tue, 14 Jul 2026 03:09:31 +0000

Week 29 of 2026 in AI dev tools: OpenAI ships an official plugin that runs Codex from inside Claude Code, a viral teardown shows Claude Code burning 33k tokens before it reads your prompt, Stably's orca gives you a control room for a fleet of parallel agents, Microsoft's Flint lets agents draw charts instead of dumping tables, Z.ai's GLM-5.2 tops Hugging Face trending, and Tencent open-sources CubeSandbox so your agents stop running rm -rf on the host.

Read the full version with code samples, diagrams, and architecture details: This Week in AI Dev: The Agent Fleet Grows Up (Week 29 of 2026)

More engineering notes: rohitraj.tech/en/notes

DeepSeek V4 API Migration Guide: What Breaks on July 24, 2026 (and the 10-Minute Fix)

Rohit Raj — Mon, 13 Jul 2026 04:08:10 +0000

On July 24, 2026 at 15:59 UTC, DeepSeek fully retires the deepseek-chat and deepseek-reasoner model names — every API call still using them starts returning errors. The replacement names (deepseek-v4-flash, deepseek-v4-pro) take ten minutes to wire in, but two silent gotchas can wreck your bill or your latency: thinking mode moved from a model name to a request parameter, and the naive migration path can turn your cheapest endpoint into a reasoning-token furnace. Here is the exact before/after code, the Flash vs Pro decision table, the Anthropic-SDK routing trick, and how I would stage the cutover in production.

Read the full version with code samples, diagrams, and architecture details: DeepSeek V4 API Migration Guide: What Breaks on July 24, 2026 (and the 10-Minute Fix)

More engineering notes: rohitraj.tech/en/notes

Stop Your AI Coding Agent Running rm -rf: Command Guardrails Compared (2026)

Rohit Raj — Sun, 12 Jul 2026 11:25:07 +0000

Destructive Command Guard (dcg) trended on GitHub in July 2026 (Rust, MIT, 2.3k stars) as a sub-millisecond PreToolUse hook that blocks your AI coding agent from running rm -rf, git reset --hard, force pushes and DROP TABLE before they execute. It wires into Claude Code, Cursor, Codex and Copilot in one install. But Adversa AI's GuardFall research bypassed the command guards in 10 of 11 popular agents. This is the builder's read: how dcg works, how to install it, whether these guards actually hold, how dcg stacks up against agent-guardrails, Shellfirm and SigmaShake, and exactly how I'd wire real agent safety into a production workflow — guard plus sandbox, not guard alone.

Read the full version with code samples, diagrams, and architecture details: Stop Your AI Coding Agent Running rm -rf: Command Guardrails Compared (2026)

More engineering notes: rohitraj.tech/en/notes