I'm interested in adding WebGPU support for a distributed ML project, so I wondered how much faster it is than CPU. I asked Claude to build this, then add export functionality. You can use it to see how much faster inference can be using WebGPU on your system. You can download the results as a .csv file if you'd like to save them.
Hi HN,I've been working on redress, a failure-policy library for Python services. It treats retries, circuit breakers, and stop conditions as coordinated responses to classified failure, rather than independent wrappers. The goal is to make failure behavior explicit, bounded, and observable across a codebase.I kept running into the same problem in distributed systems and data pipelines I worked on. Retry logic grows organically at call sites and circuit breakers live somewhere else entirely
Hi HN,I’m sharing the alpha release of S2C, a state machine replication system built atop S3.The goal is to enable a distributed application to maintain consistent state without needing a quorum of nodes for availability or consistency.The idea came from a side project that was using S3 and where I needed strongly consistent distributed state but wanted to avoid adding a separate consensus dependency. I initially tried to use S3 directly for coordination, but it became messy. Eventually, I reali
Today, we perform calculations in the way, that is easiest for the human mind. Maximum number we use, is the number of our fingers. When I examined how computers actually calculate, I realized what we call “thinking”. Computers perform only 0-1 calculations, but we can use it’s fast memory: to remember results of many simple calculations, and we sum them up. Then, computer „remembers” results of much more digits operations and sums them up correctly. See, whenever a computer performs a calculati
I built workflow because I was tired of the "infrastructure tax" required to run simple task pipelines. Most orchestrators assume you want a distributed system with a web UI and a database server.I wanted something simple and reliable, but with the features of a modern data orchestrator:- Stateful: It uses a local SQLite DB to track every run and retry. If a task fails, you can fix it and resume.- Deterministic: Strict topological ordering based on depends_on logic.- Static: A single G
<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Budget_with_Gemini_hero.max-600x600.format-webp.webp">Learn how to use simple Gemini prompts to create a 2026 budget, find hidden savings and organize your spending.
Hi HN,I've been building a small open-source CLI called Lok for orchestrating multiple LLMs in a single workflow.Instead of a single "agent", Lok treats models as interchangeable tools: one can plan, another can execute or critique, and others can verify output. The focus is on inspectable, composable workflows rather than autonomous behavior.It's local-first, CLI-driven, and intentionally small. I've been dogfooding it for real development work and documenting what work
AI coding agents read specs and start building — but nothing in the CI pipeline validates that those specs are compliant, complete, or that requirements survive the iteration process.nod is a lightweight GitHub Action that scans project documentation against community-maintained compliance rule sets. It catches missing controls, regulatory anti-patterns, and requirement drift before or during agentic development. It also includes an Agentic Action Security pack for validating agent guardrails an
AI coding tools excel at generation but fail at organization. Most demos built in 10 minutes become unmaintainable in 6 months. This framework addresses that gap. It provides: • Tiered complexity model (Lite/Standard/Enterprise) - match structure to project scale • Skills + Workflows architecture - reusable automation across projects • AI-optimized documentation (GEMINI.md) - reduce iteration cycles • Consistent patterns - same structure across all projects The focus isn't "
Kling just announced VIDEO 3.0 - a significant upgrade from their 2.6 and O1 models.Key improvements:*Extended duration:* • Up to 15 seconds of continuous video (vs previous 5-10 seconds) • Flexible duration ranging from 3-15 seconds • Better for complex action sequences and scene development*Unified multimodal approach:* • Integrates text-to-video, image-to-video, reference-to-video • Video modification and transformation in one model • Native audio generation (synchronized with video)*Two vari
Hi HN,I’m working on AI agents used for software development. These agents automatically spin up short-lived app instances – for example per pull request, per task, or per experiment – each with its own temporary URL.Auth is handled in the standard way:- OAuth2 / OIDC- external identity provider- redirect URLs must be registered in advance and be staticThis clashes badly with short-lived apps:- URLs are dynamic and unpredictable- redirect URLs can’t realistically be pre-registered- auth bec
Hi HN — I built Terminal MCP, a utility that exposes terminal sessions over MCP, primarily to help LLMs interact with CLIs and TUIs during development and debugging.It’s also useful outside of AI: it provides a general way to run terminal workflows through a controlled interface, with better isolation and observability than a raw shell.Recent addition: Sandbox mode, based on Anthropic’s work on safe tool execution. It lets you run sessions with explicit controls over what the process can access
I was looking for a way to build projects and ideas in the background while I was off doing something else. I felt like coding agents by themselves could do a certain granularity of work, but I wanted to try and push it further. So I built Orrery.What it does:- Take an idea or spec and produce an implementable plan (steps, dependencies, outputs)- Refine, simulate, and review the plan in a trackable way- Execute the plan with a deterministic step graph (same plan gives same execution order), with
My primary workflow now involves coordinating multiple, independent Claude instances to handle different parts of a codebase simultaneously.The biggest friction point I found was context switching. I tried manual git worktree management, custom Claude skills, and various wrappers, but they all felt too heavy, slow, or restrictive. I wanted something that solved the overhead of worktree management without trying to "own" my entire development process.I built cwt to bridge that gap. It i
About a year ago, I was building DockFlow (a macOS app for managing macOS dock presets) When I bumped into ExtraDock on Reddit. The original creator had built a cool tool: create multiple floating docks on your Mac, position them on different monitors. But the app was hard to maintain, and the developer was looking to move on.I loved the concept. I saw how ExtraDock and DockFlow could work beautifully together. DockFlow manages your dock configurations, ExtraDock gives you multiple docks pe
Distributed Systems - Architectural Model client server #distributedsystems #computersciencecourses #computerscience ...
In this stream we work through the fly.io distributed systems challenges (https://fly.io/dist-sys/) in Rust, and solve all the way up to ...
Accompanying lecture notes: https://www.cl.cam.ac.uk/teaching/2122/ConcDisSys/dist-sys-notes.pdf Full lecture series: ...
This video discusses what git cherry-picking is and how it is used in an advanced workflow. This video is intended for sw ...
Listen to the full episode here:On the podcast today we have Roman Axelrod, a web developer from Israel. Roman joins us to ...