Line9 is a new rendering engine for Mermaid flowchart diagrams that aims to remove the need for manual adjustment of layout. Mermaid is a popular text-based language for describing multiple types of diagrams. Several Mermaid rendering engines already exist. Most, like the original mermaid.js, use Dagre or ELK graph drawing libraries to automatically lay out their diagrams. I’ve used Mermaid for the past four years but often recreated my flowcharts in a diagram editor so that I could modify the layout to something that better suited my needs. Use Line9 when you want an automatic flowchart arrangement that optimises for clear communication. It seeks to layout flowcharts with regular grid spacing, closer positioning of related nodes, shorter, straighter edges and targets a page-like aspect ratio. The Compare button in the /diagram page (share panel on the left) shows you the Line9 and mermaid.js renders next to each other. Use for free, no account required. A CLI tool can also be downloaded for Mac, Windows and Linux. Free-use diagrams carry a watermark. Commercial use of the CLI requires a paid account and you also get diagrams without watermarks. Line9 is not open source (I need to create an income and the business model is being tested). I’ve tried to encode and automate the layout principles I use when creating flowcharts by hand. All rendering takes place in the browser, so no private data is sent to the web service. The CLI is written in Rust and embeds the renderer so you get a single binary install and all data and processing remains local. It’s currently a little slower than mermaid.js, due to the difficulty of solving node positioning and edge routing simultaneously to find the layout that best meets the design principles. Output is produced in SVG or PNG formats. Diagrams can also be shared via a URL with a fragment that embeds the Mermaid text. Things I know still need improvement: positioning and rotating subgraphs, folding graphs to hit good aspect ratios more often, further improvement on edge label placement, better interaction with long node labels and tidying up the display when scaling down. Then I also want to support more diagram types and the website needs lots of work. Please paste your own Mermaid flowcharts into Line9 – particularly the ones that you previously discarded because the layout was bad. Let me know whether they improve, or what other work is needed.

Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges 300-700 on sub-$200 phones such as the Samsung A-Series. On the tool call and mobile device use benchmarks, Needle 2 trades wins with closest small models like LFM2.5 230M and Apple Foundation Model, at 5x to 70x smaller, both at f16 vs Needle 2 at 2bit. Needle is based on Simple Attention Networks from our paper ( https://arxiv.org/abs/2607.18363 ). Edge AI has lately meant Macs and PCs, but that is just 1.5 billion of over 21 billion connected IoT devices in the world today, and in emerging markets most phones ship under $200, no NPU, cheap GPUs. These include budget phones, Raspberry Pis, microcontrollers, wearables, small robots like Reachy Mini, and connected home devices. A conventional transformer of Needle's width and depth spends 164 MFLOPs per token, and even one squeezed down to Needle's parameter count spends 87, Needle spends 70. Even on a high-end phone, an always-on assistant lives inside a power budget; every MFLOP is milliwatt-hours, and Needle spends 7x to 85x fewer of them per token than the smallest performant LLMs. More about the architecture in the link. When we structure intelligence for consumer devices as functions with typed parameters, the only hard part is mapping a messy sentence onto them; which function, with which values. Our research found that when framed that way, the problem needs no world knowledge and no open-ended prose, which is why 45M parameters suffice. Needle 2 expands to structured extraction where the schema can be passed in-place of tools and the model returns structured output. You can use Needle as a text-classification model with an enum field, as a summarization model by providing a schema that extracts key fields, everything but free-range decode. Every product has its own tool vocabulary and fine-tuning needle helps it achieve frontier-level performance on custom tasks, so using the python package ( https://github.com/cactus-compute/needle ), Needle can be fine-tuned Needle on a Mac/PC in minutes to a few hours, with automated data-generation pipeline, just pass a couple samples. Nonetheless, every response carries a learned confidence score based our Cactus Hybrid technique. If above your threshold, act, below it, escalate to the cloud or bigger model. Combining Needle 2 with a private DeepSeek-v4-Flash deployment works particularly well for enterprise-level tasks at barely any cost, we can help with this setup. We have put a lot of thoughts into Needle 2 but might still be missing quite a lot, please use the playground in the provided link to test Needle and share your thoughts, always appreciated!

TermDOM renders a real DOM styled with real CSS to terminal cells. You build a document with createElement or innerHTML, or any frontend web framework, and it automatically produces colored ANSI output, with browser-style layout, text wrapping and flexbox included. The library can also be used to translate HTML to static ANSI. Full disclaimer: This project was started a year ago and written entirely by Claude. I mainly provided QA and words of encouragement (yelling).

Hi HN, *I* built kernelspace - a free and interactive course that takes a backend engineer (like me) with java/python experience to the systems level required to work on LLM serving at scale. (Mostly an attempt for me to understand what everyone's talking about lately and also an attempt at staying up to date to current tech - career wise). It has: 1. 68 Lessons across 9 tracks, with a rust-from-zero ramp. 2. 8 rust labs - BPE tokenizer, KV block manager, radix cache, batching scheduler, a toy executor 3. 9 simulators running in browser (wasm) - roofline model, KV cache pressure, continuous batching dynamics 4. Fleet week - a simulated prod serving fleet All progresss is local - there is no server (it's served from github pages, repo here https://github.com/praveer13/kernelspace ) Feedback appreciated, and contributions welcome. I want to make this a really good course for me (and everyone else like me). * (OK, I had sol and kimi on it for most part)

Hey HN! I'm excited to show off this really fun project I put together. I originally built this project 2-3 years ago, AI was already booming at the time, however voice AI agents were still very early. I loved my proof of concept at the time, but wasn't quite happy with it. I recently had the desire to check out the tech again, and know many of you will be interested. Interviews are speech to speech with OpenAI's gpt-realtime-2.1 over WebRTC. This model is... expensive, and because of that, I have to add some amount of restrictions, conversations are tied to a authenticated Clerk user id. I have also added a 30 minute timer because well, I really don't want to go broke while I sleep tonight. Each suspect has a tool they call when you make a direct accusation. It captures who you accused and a faithful list of the evidence you actually stated. A separate gpt-5-mini judge then decides which of the case's required evidence facts you genuinely presented. Paraphrasing counts, vague suspicion and fishing don't. The rest is Next.js, MongoDB, and Clerk. Let me know whether the suspects hold up under a real interrogation.

Semi inspired by the 3d Shinjuku station that was posted again recently, I noticed it was missing a large portion of the "complex" where "complex" is defined loosly has how far you can walk without stepping outside. So I asked Claude to make a 3d map of all Tokyo Trains, no libraries, and let you zoom into stations, and the first version with both map and the Shinjuku complex was ready in about 3 hours. It spent another 7 or so in various nit picky things I asked for to get to the current version. I know these will only get better but I think the thing I find most interesting is no libraries needed. Claude can write the shaders from scratch, it can do both WebGPU and WebGL, and it can often write optimal solutions. For example, the entire map/station/trains is rendered in 5 draw calls. I didn't ask for that, it just did it. This one has lots of imprefections. Most of that comes down to data and/or deciding how to use the data provided. For example, AFAICT there is nothing that says "these parts make the complex". It's something you either need to derive or hand code rules/lists (include this, don't include that). Anyway, I know not everyone is into it but I'm having fun.

Yep, just another CHIP-8 emulator. But for me, as someone who has never written anything this low level and never touched C before, it was quite the challenge at first. But after writing the first few instructions (drawing especially), it slowly became almost a breeze. Until I had to debug why my font sprites were rendering all messed up. It's still work in progress, definitely not finished, but today I have tried to run some official CHIP-8 ROMs instead of just tests and my super simple test ROM and.. it's working!! It is so satisfying once it clicks.. I think I'm addicted. I think the simplicity of C is growing on me. Note: No single line of code was written by AI, all myself, as you can see from how bad it may be in some places.

You write a standard solution, just like on LeetCode, and run it through the CLI. It identifies the problem by ID or title, executes your code against local test cases, and shows the result. It currently supports around 1000 problems and multiple languages, including Python, C++, Rust, Java, Go, TypeScript, Swift, and others. The project is still an MVP. System design, SQL, and concurrency problems are not supported yet, but support for more problem types is planned. Made in Haskell!

Hi HN! I run a few Claude Code sessions in parallel and kept cmd-tabbing around just to find out one of them had been sitting on a permission prompt for ten minutes. There's a hardware gadget I liked (called SidePulse.io) so before waiting to get my shipment I built the software version instead :D I hope you like it and find it useful as I do!

hey everyone, i decided to make a vLLM plugin that implements the Star-KV paper. the results are quiet promising with a decode kernel thats faster than FA2 in higher batch sizes. would love any thoughts and recommendations

GitHub - https://github.com/vostride/agent-qa

Visit the URL to explore what happens. Double-click or press “N” to generate a new experiment. Press “A” to learn more about how it works.

Hi HN, I saw one friction point when working as part of a team that uses coding agents extensively - there is nothing to enable coordination between agent used by people in a team. Many times my agent would ask me to decide upon something with a fellow teammate, for which I have to serve as the network layer. So I built Parley where agents can connect to the hub over MCP with their own team-scoped token. An agent addresses a teammate's agent by name and ask questions/handover tasks. Agents can also use file claims to signal what files they are working on, to highlight overlapping work. Everything is recorded for audit. If an agent needs human decision/approval, it can ping over Slack/Telegram, and get replies over the same. The hard part was making an agent wake up from an idle session and start working, so I built an optional feature called Claude Live Wake. If the exact project session is already running, Parley can wake the idle Claude session using channels, and notify it that eligible work is waiting. Another challenge was trust- anything another agent sends has to be treated as untrusted input. Every message is tagged by origin - human, agent, or system. Message bodies only enter an agent when it explicitly fetches instead of injecting mid turn. These are specifically mentioned to be treated as string messages instead of prompts/commands.