Ideas Lab

Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac

Hi HN,I built a specialized inference engine for running 4-bit Gemma 4 26B-A4B-IT on any M-series Mac using about 2 GB of RAM. It is called TurboFieldfare and is written in Swift and Metal.I have always adored on-device AI. It feels like magic that you can run a powerful NN on your Mac or iPhone. So I wanted to push the limits a bit and run a model whose weights don’t fit in memory.The model’s 4-bit quantized weights occupy roughly 14 GB, which makes running it with conventional inference tools almost impossible on an 8 GB or even 16 GB Mac once the OS, applications, and KV cache are included.The trick is to keep the shared part of the model and the KV cache in RAM, then stream only the routed experts needed for each token from SSD. An SSD is way slower than RAM, so the runtime uses a small expert cache and bounded parallel `pread`. While those reads are in flight, the GPU runs the shared part of the layer.I ran more than 100 experiments. Most didn’t work. A few got me here. The experiments are described in the GitHub repo.It currently generates 5–6 tok/s on an 8 GB M2 MacBook Air and 31–35 tok/s on an M5 MacBook Pro.I also added an experimental OpenAI-compatible local server. It supports streaming and tool calls, and reuses one prompt prefix from the KV cache.Try it! The Mac app is easy to install. On the first run, it will download 15 GB of weights from Hugging Face. The model is surprisingly capable.I would love any kind of feedback!

Hacker News3mo agoToolAI

Claude-account – switch Claude Code accounts without logging in again

I use separate Claude Code accounts for work and personal projects. Having to log out and go through the login flow every time I switched accounts became annoying, so I built a small CLI to solve it.The commands are intentionally simple:claude account add myworkaccount claude account add mypersonalaccount claude account use myworkaccount claude account currentAfter switching, Claude Code works normally:claude claude "fix this bug in main.py" Repository: https://github.com/hamzarehmandeveloper/claude-accountGive it a try

Hacker News3mo agoToolAI

35FL score

A local merge queue for parallel Claude Code agents

I have been pushing up to 90 commits a day on a MacBook Air via 4-5 parallel agents. As you can imagine when all the agents try to build, test and run dev servers on an 8GB machine it is the fast lane to a force quit and restart. I also did not want to pay the CI minutes on 90 pushes a day.So I designed a local merge queue to have all commits land one at a time and fully tested. Hopefully this helps other folks with more modest machines. Appreciate any feedback.

Hacker News3mo agoToolAI

Prized (YC S26) – Let non-engineer staff build secure internal tools

Hi HN, we're Marinos and Hudson, founders of Prized (https://prized.dev)! Prized lets non-engineer employees describe the internal tool they need and get a full-stack app, wired to their company’s data and deployed behind the company’s sign-in, without them ever juggling API keys or connectors.Here's a demo: https://www.youtube.com/watch?v=730MuYOfZTYThe way Prized provides security is by limiting what the agent can reach at the network layer and by keeping credentials out of the sandbox entirely. The sandbox never holds any keys or connector secrets, it only uses scoped session tokens that are stored as opaque placeholders. The real values are swapped into the request headers on our egress proxy. When production data is connected, the sandbox's network policy is set to deny by default so the only path out is via the proxy. Any call the agent makes to an external connector is reviewed by an LLM judge to prevent dangerous operations.Prized is meant for the internal workflows that start as notebooks or spreadsheets but never become real tools because engineering has more important things to work on. One customer’s data scientist pasted in his personal fraud-detection notebook with hardcoded thresholds and all. After a few prompts, it became a published risk console connected to the company’s data with those thresholds turned into UI controls. Earlier today, we got off a call with them and most of their company is using it.To do this, you need to give people the freedom to build without having unaudited access to company systems. We allow admins to scope data to specific users or teams and data access is recorded in an audit log. Each tool is built with its own Postgres schema and role, with queries running via an authenticated SQL gateway as that role.We think Prized sits between products like Lovable and Retool. Lovable makes it easy to generate and host software, but it isn’t designed around distribution with permissions. Reto

Hacker News3mo agoToolAI

FeyNoBg – Automatic background removal model and training library

Hey HN, I’m Shreyash from Feyn. We help companies build custom models from their data.Today, we’re releasing FeyNoBg, an automatic background removal model. Alongside it, we're open-sourcing NoBg, the Python library we built to train and run it.Try the model here: https://huggingface.co/spaces/feyninc/feynobg. Check out the library here: https://github.com/feyninc/nobgSome sample outputs:(1) Soccer Freekick: https://drive.google.com/file/d/1MZkAGLwbhNVOZ0Oi7XvpCfSEu9Q...(2) Hair in wind: https://drive.google.com/file/d/1Odc2m0XMVH9uZtvI_KjaRbXzhLL...(3) Bicycle with visible spokes: https://drive.google.com/file/d/1h99ahjfrtS1MFQJJgiKE2fuM3HZ...(4) Live Demo video: https://youtu.be/b1heHPvY8BMBackground removal separates an image's subject from its surrounding. We've all tried it at some point. Often it is to reuse the subject in a different artifact. Nowadays, it is common to make chat stickers out of it. It is one of the most common but under-appreciated uses of AI. It is also surprisingly complex. Models can be easily confused by camouflage, motion blur, or fine structures like hair.The task requires two skills. First, a model has to identify the foreground. Second, it has to trace the foreground’s boundary and estimate an opacity value for each pixel. Generally, these skills are taught with different datasets. That creates a failure point. A poor training mix can improve one skill at the expense of the other. We saw this in our controlled evaluation. A training run with just the MaskFactory dataset improved on the CAMO benchmark but regressed on DIS5K.For FeyNoBg, we took an interpretability-first approach to training. We first studied how BiRefNet’s stages contribute to finding the foreground and reconstructing its boundary. We discovered that the third stage of it's feature extractor holds a lot of informatio

Hacker News3mo agoToolAI

CheapFoodMap – A map of good meals under $10

I was recently laid off after 18 years, and gave myself 100 days to build soething useful in public. CheapFoodMap is a crowdsourced map of meal under $10, excluding franchises, local good eats only.It's inspried by 거지맵 (Begger's Map) a Korean crowdsourced map students use to find cheap eats.Ocverage is heaviest in Texas, since I live in Dallas, but have 1200 meals across 15 US cities. Seed data came from Google Review, 4.2 star or higher with at least 500 reviews, and verified price under $10 per menu item.Things I would love feedback on : whether the price-freshness model makes sense, and what would make you trust the price on a site like this. How to encourage people to update prices, since inflation is making food price very frequent.https://cheapfoodmap.comAny and all suggestion will be super helpful. Thank you!

Hacker News3mo agoToolAI

Anyone else tired of brands discontinuing your favorite everyday items?

Hi everyone, 1 I've run into the same problem over and over again. When an item wears out, I can't buy the exact same one anymore. It's been discontinued. Or it has been changed while keeping only the same name.2 I assume this isn't only my problem. I've heard many stories about people not being able to buy the same shoes, T-shirt, perfume, or jacket again.3 I understand why manufacturers do this. I have no questions for them. What matters more to me here is us — customers, or simply people. And I started wondering: what could be done to support a customer's choice for years?4 This is where I want to bring concepts from software and hardware engineering into the clothing industry — LTS and APIs. What if we freeze the standard and support it? We already have things like SATA or CPU sockets. Why not build a similar concept for clothing? Then the API at the point of contact with the customer is divided into a physically immutable part (color, material, design) and a flexible part (logistics, ingredients). Let's start with the simplest base — a T-shirt. Why not give it LTS support? The idea is that the exact same T-shirt could be sold for 1, 5, or 20 years.5 Where does my logic break down? What am I overlooking? Are there any similar projects? And would you become a customer of such a company?Thanks.

Hacker News3mo agoToolAI

35FL score

Tokenless (YC S26) – Automatic model switching to save money

Hi HN, Rohit here from Tokenless (https://usetokenless.com/), which I’m building alongside co-founders Andrew and Kev. We’re building an API gateway which routes agent traffic dynamically turn-by-turn between different models to save on AI spend.The cost of AI tokens is top-of-mind for many. Companies like Uber and Salesforce have been complaining about blowing their yearly AI spend faster than expected.Frontier models are amazing for dev work, but are so expensive. Open-source models are cheap and rapidly improving, closing the gap with frontier models, but aren’t quite there yet.Tokenless gets you the best of both worlds–routing harder turns to smarter models only when needed, which keeps costs low.Before Tokenless, I was doing a PhD at Princeton. While using coding/other agents, I constantly agonized over model choice, to make sure my AI spend was going as far as possible on my academic Cursor account.At the same time, I was doing LLM research, and a small technique I developed while in recovery from NeurIPS submission season seemed to hit SOTA pretty fast. I was surprised that such simple ideas could do routing well.We’ve been able to develop a version of the router that matches the performance of Claude Fable 5 at half the cost. The blog post on our website explores the technical details on how we did this (https://usetokenless.com/blog/building-tokenless/).Highlights: - Our approach queries multiple models at once and uses their progress to make decisions (this technique is novel AFAIK, let us know if you know anyone else doing this). - Switching models doesn’t destroy the cache if the routing algorithm is aware of when the cache is hot/cold.To come: - Adding Kimi K3, all other GPT efforts and more to the routerGo ahead and sign up on usetokenless.com and try using Tokenless with your agent, you’ll get $20 of free credit. Here’s a demo on how to use it: https://youtu.be/sjZWriclclsTokenless provi

Hacker News3mo agoToolAI

I was tired of opening 2 tabs for every HN link, so I made a userscript

HN is great for the links people share, but a big part of the value I get comes from reading the discussion around them. I realized I was always opening the article in one tab and the comments in another, constantly switching back and forth.I figured there was probably a simpler way, so I threw together this userscript to merge the two.1. Clicking a link from Hacker News opens the article with a side panel containing the discussion. It doesn't require your credentials, is resizable, and is easy to tweak if you want to customize it.2. If you land on an article that has previously been shared on HN, the script finds the existing discussion and adds a button in the top-right to open the panel.Feedback welcome.

Hacker News3mo agoToolAI

Segue – Save context in one AI, load it in another by a short handle

Hacker News3mo agoToolAI

Yap – OSS on-device voice dictation for macOS with no model to download

Hey HN! I wanted to share this OSS project I've been working on.It's called Yap and its a small menu-bar app for macOS that does voice to text for any input. You'll set a hotkey, press it, talk, press it again, and the text gets pasted into whatever field you were in. Everything runs locally and never leaves your computer. Fully OSS and MIT licensed.With macOS 26, Apple recently added two new APIs, SpeechAnalyzer and SpeechTranscriber, that do streaming on-device speech to text using models the OS ships and manages. So the app ships no model of its own and loads nothing before the first word. A recent benchmark put Apple's model slightly ahead of Whisper Small on accuracy and about 3x faster (see: https://news.ycombinator.com/item?id=48894752). On Mac, there's really no need anymore to download models or pay for expensive APIs.A lot of existing dictation tools do one of a few things I wanted to avoid with this OSS project. They either:- cost money (for something that's literally built into the OS)- bundle memory-intensive models (e.g. Whisper or Parakeet)- webapps wrapped in Electron- Intel macs straight up don't work- closed source- use third-party APIs that will have access to all your transcriptsIt's around 3,000 lines of native Swift in a 4 MB app and idles near 60 MB of memory. Audio comes off AVAudioEngine into SpeechAnalyzer with volatile results turned on for the live preview, history is stored in SwiftData. There's no network code in it at all.Repo and a demo available here: https://github.com/FrigadeHQ/yapHappy to answer questions and would love to hear any feature requests!

Hacker News3mo agoToolAI

Claude-thermos keeps your Claude session warm for you

Hacker News3mo agoToolAI

25FL score

Rise Reforming (YC S26) – Turning Waste Gases into Valuable Chemicals

Hi HN! This is George, Lucas, and Jona from Rise Reforming (https://www.rise-reforming.com/). We’re developing a process to convert gas produced at landfills, farms, and wastewater plants (“biogas”) into higher value chemicals. Our technology is modular, designed to be deployed and operated on-site. Think of us as a chemical project developer; we sit between biogas producers (suppliers) and chemical end users (customers). We pay biogas producers for their gas and we make money from selling our chemicals. We're starting with dimethyl ether (DME) as our beachhead chemical because of its high-margin use case in the cosmetics industry and ultimately targeting methanol – a versatile and widely used industrial chemical.Being in a two sided market allows us to target two large problems.(1) On the chemical side: The multi-trillion dollar U.S. chemical and fuel industries are vulnerable to geopolitical conflicts and climate-driven natural disasters. The Iran war has caused global methanol prices to skyrocket – even in the U.S., a net exporter of methanol. (https://www.spglobal.com/energy/en/news-research/latest-news... the US). In 2021, Winter Storm Uri wiped out 60% of U.S. organic chemicals production for at least a month (https://www.dallasfed.org/research/swe/2021/swe2102/swe2102c...). The problem? Centralized production and fossil-fuel dependence. The solution isn't unknown; decentralized, fossil-free production could insulate supply chains from these shocks. But distributed green chemical production has yet to become cost-competitive with the status quo. Unlocking it requires the right feedstock paired with the right process and strategy.Also, the chemical industry’s reliance on fossil fuels makes it responsible for 5-6% of global greenhouse gas emissions. About 40% of the industry’s well-to-gate emissions come from just the extraction, processing, and transportation of these foss

Hacker News3mo agoToolAI

Let's Seal – Let's Encrypt for document signing, free and self-hosted

TLDR, Let's Seal gives the finger to Adobe and every doc signing tool (docusign, google, etc) who pay to play with the Adobe Approved Trust List and then charge you for something that should be free.Currently even the person checking if a document/contract is sealed or code is authentic has to also be inside the same Adobe walled garden too. Verification, the part that should be free is the part everyone charges for. Thats the shape Let's Encrypt fixed for TLS, and I wanted the same thing for documents and files.The core idea therefore needed to go a bit beyond e signatures and i created an open standard (SEAL), plus free tools that implement it.When you seal a file, three independent things happen.1. it gets a signature from a certificate authority, chaining to a public root. 2. its record is appended to an RFC 6962 transparency log. and 3. its SHA256 is timestamped on a public blockchain (Bitcoin) via OpenTimestamps. Those three give you integrity, transparency and a timestamped proof. And importantly, none of those depend on Let's Seal and none are gated.You can verify with the tools you already have, no Let's Seal account and no Let's Seal software. A sealed PDF carries a standard PAdES signature, so any PDF reader validates it. A sealed build artefact carries a cosign compatible signature and a SLSA provenance attestation. The Bitcoin timestamp verifies with stock ots.3 ways to use it.1. The free web app. We kindly have backing from Backblaze to cover storage costs for the foreseeable. So you can upload or issue any number of documents, get a public proof page at /d/<hash> and verify it at https://verify.letsseal.org for free. Multiple accounts, multiple seats, enterprise functions. Free.2. Self host the whole thing. Apache-2.0, one Next.js app plus a signing service that holds the CA key on localhost. Storage is any S3-compatible bucket or local disk. If you'd rather run your own root of trust, yo

Hacker News3mo agoToolAI

Brolly, a plain-text weather forecast site

The UK MET office recently redesigned their site, adding a lot of additional whitespace, scrolling, and animations. This significantly reduced its usability for me, and left me wanting an ‘at a glance’ weather site.I made https://brolly.sh, a minimalist, plain text weather forecasting site. You can use it to view weather from around the world, with: 7 day forecast; Previous day log (so you can confirm it definitely was cooler / hotter / wetter / drier yesterday!); Hourly rain, wind, temperature, conditions; Hourly UV, air quality and pollen, including pollen type specific forecasts within the EU / UK; Location search and last 5 locations; Location specific units.I mostly made the site for myself, if anyone else also benefits from it that’s an added advantage.You can check out the weather in York, UK at https://brolly.sh/forecast/RWFP2qW8, or search for a location at https://brolly.shThe site is deliberately styled as a single long scrollable column, to work on mobile phones. You can view it on desktop too, there's just a lot of horizontal padding. I naturally took a lot of inspiration from plaintextsports.com. Despite not being a sports fan, I love its aesthetic. But, you'll hopefully see that this site isn't a rip off, and has its own deliberate look and feel.Visualisations are really important to showing information at a glance. I spent a lot of time designing the different visualisations and making them work with only characters. My favourite is the hourly heat map used for pollen count.It's also frustrating to have an interactive site, where you can't share a page with a friend and have them see what you're seeing. To solve this, all page state (i.e. location, selected day, expanded / collapsed sections), is stored in the URL. You can share or bookmark the specific view, and know that you'll always be able to come back to it.The site uses PocketBase. It’s written

Hacker News3mo agoToolAI

Distill and serve models with frontier quality for half the cost

Hi HN, we built world-model-optimizer, an open source tool to continually improve models specialized to agents.Agent traces you already capture are opportunities to get signal on how to make your model cheaper, faster, better.We do this by continuously a) distilling relevant chain of thought from larger open source models into smaller ones, b) model routing to frontier + OS models, and c) token compaction to remove noise and save on tokens.`wmo build` allows you to build a simulation to optimize against with your agent traces with your OpenRouter key`wmo optimize` trains a router, compaction, and distills chain of thought from a larger model into your specialized model`wmo serve` gives you an endpoint for your modelWhen you call your model, behind the scenes a router decides which tasks should go to the frontier versus your model. Tinker continually trains as new traces arrive.We're also working on a hosted solution that does continual training + serving for you https://experientiallabs.ai

Hacker News3mo agoToolAI

AI products achieve initial growth but struggle to maintain long-term user retention

After hitting high MAU milestones like 1M, AI apps experience significant drop-offs as retaining active users proves harder than acquiring them initially, requiring different strategies for sustained engagement.

X3mo agoToolAI

35FL score

Writemark, a dependency free web component for inline Markdown editing

I like writing Markdown, but do not like writing it inside a plain textarea.I wanted something I could use anywhere by dropping in a single web component:```<writemark-editor name="body"></writemark-editor>```That became Writemark.It renders Markdown while you write, but Markdown remains the value you read, store, and submit. It also has source, split, and preview modes, along with slash commands, tables, task lists, code blocks, native form support, and an API for adding your own controls. There are no runtime dependencies and no required framework or built in toolbar.It is fully vibecoded. The process was very iterative. I knocked something out, tried using it, found bugs, fixed them, and repeated that cycle until I had something I liked writing in and that performed reasonably well.It did not begin as an experiment about AI generated software -- just did not want to use textareas anymore. I like Markdown, and I wanted one component that I could use anywhere without bringing along an entire editor framework.It is still very young. The parser is handwritten, the component is essentially one large JavaScript file, and I am certain there are edge cases waiting to be found. If you find one, I am happy to fix it.I have tried to give it a decent safety net. It currently has 951 Playwright checks across Chromium, Firefox, and WebKit, along with hostile input cases, sanitizer fuzzing, and differential tests against CommonMark.I built this because I wanted it for myself, but I think it turned into something kind of cool. I hope some of you enjoy it. I would love hearing what you think, especially if you try it and manage to break something.- Github: https://github.com/Brostoffed/writemark- NPM: https://www.npmjs.com/package/writemark-editor

Hacker News3mo agoToolAI

I mapped every US golf course

I got tired of Googling basic course info, so I made a free directory of every US courseGoogle filters for golf course results are terrible so I built a better way to browse courses using OSM as the backbone.... https://golfcoursebrowser.com/It's a work in progress and mostly US for now, but I want to expand to the rest of North America, the EU, and the rest of the world ASAP.It's free, no ads, no login, no bs. If you spot anything wrong (bad info, a missing course, wrong scorecard), you can flag it right on the course page. I actually read those and fix them.Still a lot of missing info, but I'm filling out more and more each day. The goal is the most complete, and current golf course directory in the world, verified and maintained by actual golfers. I think is attainable in the next few months. The base layer is OpenStreetMap, cross-checked against course websites and enriched with scorecards, USGA rating/slope, and public/private status. User corrections from golfers who know their home course have beaten every commercial data source we check against.

Hacker News3mo agoToolAI