Show HN: Kanon 2 Enricher – the first hierarchical graphitization model https://bit.ly/4boHrAq

Show HN: Kanon 2 Enricher – the first hierarchical graphitization model Hey HN, This is Kanon 2 Enricher, the first hierarchical graphitization model. It represents an entirely new class of AI models designed to transform document corpora into rich, highly structured knowledge graphs. In brief, our model is capable of: - Entity extraction, classification, and linking: identifying key entities like individuals, companies, governments, locations, dates, documents, and more, and classifying and linking them together. - Hierarchical segmentation: breaking a document up into its full hierarchy, including divisions, sections, subsections, paragraphs, and so on. - Text annotation: extracting common textual elements such as headings, sigantures, tables of contents, cross-references, and the like. We built Kanon 2 Enricher from scratch. Every node, edge, and label in the Isaacus Legal Graph Schema (ILGS), which is the format it outputs to, corresponds to at least one task head in our model. In total, we built 58 different task heads jointly optimized with 70 different loss terms. Thanks to its novel architecture, unlike your typical LLM, Kanon 2 Enricher doesn't generate extractions token by token (which introduces the possibility of hallucinations) but instead directly classifies all the tokens in a document in a single shot. This makes it really fast. Because Kanon 2 Enricher's feature set is so wide, there are a myriad of applications it can be used for, from financial forensics and due diligence all the way to legal research. One of the coolest applications we've seen so far is where a Canadian government built a knowledge graph out of thousands of federal and provincial laws in order to accelerate regulatory analysis. Another cool application is something we built ourselves, a 3D interactive map of Australian High Court cases since 1903, which you can find right at the start of our announcement. Our model has already been in use for the past month, since we released it through a closed beta that included Harvey, KPMG, Clifford Chance, Clyde & Co, Alvarez & Marsal, Smokeball, and 96 other design partners. Their feedback was instrumental in improving Kanon 2 Enricher before its public release, and we're immensely thankful to each and every beta participant. We're eager to see what other developers manage to build with our model now that its out publicly. https://bit.ly/4ud0Aga March 3, 2026 at 09:55AM

Show HN: I built an AI exam prep platform for AWS certs after failing one myself https://bit.ly/4aY1AvG

Show HN: I built an AI exam prep platform for AWS certs after failing one myself Hey HN, I failed the AWS Advanced Networking Specialty exam. Studied for weeks, used the usual prep sites, thought I was ready — wasn't. The problem wasn't effort, it was the tools. Static question banks don't teach you to think through AWS architecture decisions. They teach you to pattern-match answers. That falls apart on the harder exams. So I built Knowza to fix that for myself, and then figured others probably had the same frustration. The idea: instead of a static question bank, use AI to generate questions, adapt to what you're weak on, and actually explain the reasoning behind each answer — the way a senior engineer would explain it, not a multiple choice rubric. The stack: Next.js + Amplify Gen 2 DynamoDB (direct Server Actions, no API layer) AWS Bedrock (Claude) for question generation and explanations Stripe for billing The hardest part honestly wasn't the AI — it was getting question quality consistent enough that I'd trust it for real exam prep. Still iterating on that. Early days, one person, built alongside a day job. Would love feedback from anyone who's grinded AWS certs or has thoughts on AI-generated educational content. knowza.ai https://bit.ly/3MMq0R7 March 5, 2026 at 09:27PM

Show HN: A shell-native cd-compatible directory jumper using power-law frecency https://bit.ly/4cz3be9

Show HN: A shell-native cd-compatible directory jumper using power-law frecency I have used this tool privately since 2011 to manage directory jumping. While it is conceptually similar to tools like z or zoxide, the underlying ranking model is different. It uses a power-law convolution with the time series of cd actions to calculate a history-aware "frecency" metric instead of the standard heuristic counters and multipliers. This approach moves away from point-estimates for recency. Most tools look only at the timestamp of the last visit, which can allow a "one-off" burst of activity to clobber long-term habits. By convolving a configurable history window (typically the last 1,000+ events), the score balances consistent habits against recent flukes. On performance: Despite the O(N) complexity of calculating decay for 1,000+ events, query time is ~20-30ms (Real Time) in ksh/bash, which is well below the threshold of perceived lag. I intentionally chose a Logical Path (pwd -L) model. Preserving symlink names ensures that the "Name" remains the primary searchable key. Resolving to physical paths often strips away the very keyword the user intends to use for searching. https://bit.ly/3N95WIu March 4, 2026 at 11:20AM

Show HN: DubTab – Live AI Dubbing in the Browser (Meet/YouTube/Twitch/etc.) https://bit.ly/4u1yiVL

Show HN: DubTab – Live AI Dubbing in the Browser (Meet/YouTube/Twitch/etc.) Hi HN — I’m Ethan, a solo developer. I built DubTab because I spend a lot of time in meetings and watching videos in languages I’m not fluent in, and subtitles alone don’t always keep up (especially when the speaker is fast). DubTab is a Chrome/Edge extension that listens to the audio of your current tab and gives you: 1.Live translated subtitles (optional bilingual mode) 2.Optional AI dubbing with a natural-sounding voice — so you can follow by listening, not just reading The goal is simple: make it easier to understand live audio in another language in real time, without downloading files or doing an upload-and-wait workflow. How you’d use it 1.Open a video call / livestream / lecture / any tab with audio 2.Start DubTab 3.Choose target language (and source language if you know it) 4.Use subtitles only, or turn on natural AI dubbing and adjust the audio mix (keep original, or duck it) What it’s good for 1.Following cross-language meetings/classes when you’re tired of staring at subtitles 2.Watching live content where you can’t pause/rewind constantly 3.Language learners who want bilingual captions to sanity-check meaning 4.Keeping up with live news streams on YouTube when events are unfolding in real time (e.g., breaking international updates like U.S./Iran/Israel-related developments) Link: https://bit.ly/40HBFUo I’ll be in the comments and happy to share implementation details if anyone’s curious. https://bit.ly/40HBFUo March 4, 2026 at 02:04AM

Show HN: I built a LLM human rights evaluator for HN (content vs. site behavior) https://bit.ly/4l4c4yi

Show HN: I built a LLM human rights evaluator for HN (content vs. site behavior) My health challenges limit how much I can work. I've come to think of Claude Code as an accommodation engine — not in the medical-paperwork sense, but in the literal one: it gives me the capacity to finish things that a normal work environment doesn't. Observatory was built in eight days because that kind of collaboration became possible for me. (I even used Claude Code to write this post — but am only posting what resonates with me.) Two companion posts: on the recursive methodology ( https://bit.ly/409tFeD... ) and what 806 evaluated stories reveal ( https://bit.ly/4r7k9DW... ). I built Observatory to automatically evaluate Hacker News front-page stories against all 31 provisions of the UN Universal Declaration of Human Rights — starting with HN because its human-curated front page is one of the few feeds where a story's presence signals something about quality, not just virality. It runs every minute: https://bit.ly/4aKNMpG . Claude Haiku 4.5 handles full evaluations; Llama 4 Scout and Llama 3.3 70B on Workers AI run a lighter free-tier pass. The observation that shaped the design: rights violations rarely announce themselves. An article about a company's "privacy-first approach" might appear on a site running twelve trackers. The interesting signal isn't whether an article mentions privacy — it's whether the site's infrastructure matches its words. Each evaluation runs two parallel channels. The editorial channel scores what the content says about rights: which provisions it touches, direction, evidence strength. The structural channel scores what the site infrastructure does: tracking, paywalls, accessibility, authorship disclosure, funding transparency. The divergence — SETL (Structural-Editorial Tension Level) — is often the most revealing number. "Says one thing, does another," quantified. Every evaluation separates observable facts from interpretive conclusions (the Fair Witness layer, same concept as fairwitness.bot — https://bit.ly/43DzQKs ). You get a facts-to-inferences ratio and can read exactly what evidence the model cited. If a score looks wrong, follow the chain and tell me where the inference fails. Per our evaluations across 805 stories: only 65% identify their author — one in three HN stories without a named author. 18% disclose conflicts of interest. 44% assume expert knowledge (a structural note on Article 26). Tech coverage runs nearly 10× more retrospective than prospective: past harm documented extensively; prevention discussed rarely. One story illustrates SETL best: "Half of Americans now believe that news organizations deliberately mislead them" (fortune.com, 652 HN points). Editorial: +0.30. Structural: −0.63 (paywall, tracking, no funding disclosure). SETL: 0.84. A story about why people don't trust media, from an outlet whose own infrastructure demonstrates the pattern. The structural channel for free Llama models is noisy — 86% of scores cluster on two integers. The direction I'm exploring: TQ (Transparency Quotient) — binary, countable indicators that don't need LLM interpretation (author named? sources cited? funding disclosed?). Code is open source: https://bit.ly/3MJJANP — the .claude/ directory has the cognitive architecture behind the build. Find a story whose score looks wrong, open the detail page, follow the evidence chain. The most useful feedback: where the chain reaches a defensible conclusion from defensible evidence and still gets the normative call wrong. That's the failure mode I haven't solved. My background is math and psychology (undergrad), a decade in software — enough to build this, not enough to be confident the methodology is sound. Expertise in psychometrics, NLP, or human rights scholarship especially welcome. Methodology, prompts, and a 15-story calibration set are on the About page. Thanks! https://bit.ly/4aKNMpG March 4, 2026 at 01:26AM

Show HN: Interactive WordNet Visualizer-Explore Semantic Relations as a Graph https://bit.ly/4l9DCCr

Show HN: Interactive WordNet Visualizer-Explore Semantic Relations as a Graph https://bit.ly/4l7NYTv March 3, 2026 at 10:17PM

Show HN: An Auditable Decision Engine for AI Systems https://bit.ly/4r0ct6d

Show HN: An Auditable Decision Engine for AI Systems https://bit.ly/4rKkhKt March 3, 2026 at 03:03AM

Show HN: PHP 8 disable_functions bypass PoC https://bit.ly/4coTizr

Show HN: PHP 8 disable_functions bypass PoC https://bit.ly/4ckhq6k March 3, 2026 at 02:12AM

Show HN: We filed 99 patents for deterministic AI governance(Prior Art vs. RLHF) https://bit.ly/3OHLRtr

Show HN: We filed 99 patents for deterministic AI governance(Prior Art vs. RLHF) For the last few months, we've been working on a fundamental architectural shift in how autonomous agents are governed. The current industry standard relies almost entirely on probabilistic alignment (RLHF, system prompts, constitutional training). It works until it's jailbroken or the context window overflows. A statistical disposition is not a security boundary. We've built an alternative: Deterministic Policy Gates. In our architecture, the LLM is completely stripped of execution power. It can only generate an "intent payload." That payload is passed to a process-isolated, deterministic execution environment where it is evaluated against a cryptographically hashed constraint matrix (the constitution). If it violates the matrix, it is blocked. Every decision is then logged to a Merkle-tree substrate (GitTruth) for an immutable audit trail. We filed 99 provisional patents on this architecture starting January 10, 2026. Crucially, we embedded strict humanitarian use restrictions directly into the patent claims themselves (The Peace Machine Mandate) so the IP cannot legally be used for autonomous weapons, mass surveillance, or exploitation. I wrote a full breakdown of the architecture, why probabilistic safety is a dead end, and the timeline of how we filed this before the industry published their frameworks: Read the full manifesto here: https://bit.ly/4l5y3Vx... The full patent registry is public here: https://bit.ly/4l1JNbI I'm the founder and solo inventor. Happy to answer any questions about the deterministic architecture, the Merkle-tree state persistence, or the IP strategy of embedding ethics directly into patent claims. March 2, 2026 at 11:56PM

Show HN: Open-Source Postman for MCP https://bit.ly/4l4lxG3

Show HN: Open-Source Postman for MCP https://bit.ly/40EKzC1 March 3, 2026 at 12:40AM

Show HN: Vibe Code your 3D Models https://bit.ly/4aYHwto

Show HN: Vibe Code your 3D Models Hi HN, I’m the creator of SynapsCAD, an open-source desktop application I've been building that combines an OpenSCAD code editor, a real-time 3D viewport, and an AI assistant. You can write OpenSCAD code, compile it directly to a 3D mesh, and use an LLM (OpenAI, Claude, Gemini, ...) to modify the code through natural language. Demo video: https://www.youtube.com/watch?v=cN8a5UozS5Q A bit about the architecture: - It’s built entirely in Rust. - The UI and 3D viewport are powered by Bevy 0.15 and egui. - It uses a pure-Rust compilation pipeline (openscad-rs for parsing and csgrs for constructive solid geometry rendering) so there are no external tools or WASM required. - Async AI network calls are handled by Tokio in the background to keep the Bevy render loop smooth. Disclaimer: This is a very early prototype. The OpenSCAD parser/compiler doesn't support everything perfectly yet, so you will definitely hit some rough edges if you throw complex scripts at it. I mostly just want to get this into the hands of people who tinker with CAD or Rust. I'd be super happy for any feedback, architectural critiques, or bug reports—especially if you can drop specific OpenSCAD snippets that break the compiler in the GitHub issues! GitHub (Downloads for Win/Mac/Linux): https://bit.ly/3MDl1Cd Happy to answer any questions about the tech stack or the roadmap! https://bit.ly/3MDl1Cd February 27, 2026 at 06:27PM

Show HN: Logira – eBPF runtime auditing for AI agent runs https://bit.ly/3MP5orl

Show HN: Logira – eBPF runtime auditing for AI agent runs I started using Claude Code (claude --dangerously-skip-permissions) and Codex (codex --yolo) and realized I had no reliable way to know what they actually did. The agent's own output tells you a story, but it's the agent's story. logira records exec, file, and network events at the OS level via eBPF, scoped per run. Events are saved locally in JSONL and SQLite. It ships with default detection rules for credential access, persistence changes, suspicious exec patterns, and more. Observe-only – it never blocks. https://bit.ly/4sgvLW1 https://bit.ly/4sgvLW1 March 2, 2026 at 12:25AM

Show HN: InstallerStudio – Create MSI Installers Without WiX or InstallShield https://bit.ly/4ukgsOb

Show HN: InstallerStudio – Create MSI Installers Without WiX or InstallShield Hi, I'm Paul — 25 years of enterprise Windows development. I built InstallerStudio after WiX went from free/open source to $6,500/year support and InstallShield hit $2,000+/year. Every tool in this space is either unaffordable or requires writing XML by hand. InstallerStudio is a visual MSI designer built on WinUI 3/.NET 10. No XML, no subscriptions, no external dependencies. Handles files, Windows services, registry, shortcuts, file associations, custom actions, and full installer UI. $159 this month, $199 after. 30-day free trial. It ships its own installer, built with itself. Happy to answer questions about MSI internals. https://bit.ly/3MAMtAv February 28, 2026 at 11:02PM

Show HN: OpenTimelineEngine – Shared local memory for Claude Code and codex https://bit.ly/4s8g1o6

Show HN: OpenTimelineEngine – Shared local memory for Claude Code and codex https://bit.ly/4aHDwP5 February 28, 2026 at 01:00AM

Show HN: Notemac++ – A Notepad++-inspired code editor for macOS and the web https://bit.ly/405WZCO

Show HN: Notemac++ – A Notepad++-inspired code editor for macOS and the web https://bit.ly/4u5ygMQ February 28, 2026 at 12:05AM

Show HN: Lar-JEPA – A Testbed for Orchestrating Predictive World Models https://bit.ly/3P5wkUf

Show HN: Lar-JEPA – A Testbed for Orchestrating Predictive World Models Hey HN, The current paradigm of agentic frameworks (LangChain, AutoGPT) relies on prompting LLMs and parsing conversational text strings to decide the next action. This works for simple tasks but breaks down for complex reasoning because it treats the agent's mind like a scrolling text document. As research shifts toward Joint Embedding Predictive Architectures (JEPAs) and World Models, we hit an orchestration bottleneck. JEPAs don't output text; they output abstract mathematical tensors representing a predicted environmental state. Traditional text-based frameworks crash if you try to route a NumPy array. We built Lar-JEPA as a conceptual testbed to solve this. It uses the Lár Engine,a deterministic, topological DAG ("PyTorch for Agents") to act as the execution spine. Key Features for Researchers: Mathematical Routing (No Prompting): You write deterministic Python RouterNodes that evaluate the latent tensors directly (e.g., if collision_probability > 0.85: return "REPLAN"). Native Tensor Logging: We custom-patched our AuditLogger with a TensorSafeEncoder. You can pass massive PyTorch/NumPy tensors natively through the execution graph, and it gracefully serializes them into metadata ({ "__type__": "Tensor", "shape": [1, 768] }) without crashing JSON stringifiers. System 1 / System 2 Testing: Formally measure fast-reflex execution vs. deep-simulation planning. Continuous Learning: Includes a Default Mode Network (DMN) architecture for "Sleep Cycle" memory consolidation. We've included a standalone simulation where a Lár System 2 Router analyzes a mock JEPA's numerical state prediction, mathematically detects an impending collision, vetoes the action, and replans—all without generating a single word of English text. Repo: https://bit.ly/4b909fj Would love to hear your thoughts on orchestration for non-autoregressive models. https://bit.ly/4b909fj February 27, 2026 at 03:38AM

Show HN: I Built Smart Radio That Auto-Skips Talk and Ads by Using ML https://bit.ly/4l6Exnm

Show HN: I Built Smart Radio That Auto-Skips Talk and Ads by Using ML Hi, I built TuneJourney to solve a specific annoyance: radio ads and DJ chatter. The core feature is an in-browser "AI Skip Talk" filter. The Tech: Instead of processing on a server, it uses the Web Audio API to capture the stream locally and runs a lightweight ML classification model directly in your browser. It estimates the music vs. speech probability in near real-time. If enabled, it automatically triggers a "next" command to hop to another station the moment an ad, news segment, or DJ starts talking. Features: - In-browser Inference: Entirely local and privacy-focused; no audio data ever leaves your machine. - WebGL + Point Clustering: Renders 70,000 stations across 11,000 locations smoothly. - Real-time Activity: See other users on the globe and what they are listening to in real-time. - System Integration: Full Media Key support for physical keyboard and system-level Next/Prev buttons. - Customization: Includes a talk sensitivity slider for the ML model so you can tweak the threshold. Check it out: https://bit.ly/3OBjYTQ Let me know what you think! I am interested if this project is worth further investment, building a mobile app, etc. https://bit.ly/3OBjYTQ February 27, 2026 at 01:09AM

Show HN: OrangeWalrus, an aggregator for trivia nights (and other events) in SF https://bit.ly/4tT1vCg

Show HN: OrangeWalrus, an aggregator for trivia nights (and other events) in SF Two problems I encountered personally: 1) Some buddies and I went to a trivia night late last year, only to arrive to find it cancelled (with signs still on the walls saying it happened every Tuesday, etc) 2) Sourcing ideas for fun things to do in the city on a given night, in a given neighborhood. Some sites help a ton (e.g. funcheapsf), but often don't have everything I'd want to see, so we decided to build that out a bit. Anyway, I built this originally to solve #1, then a buddy and I expanded it to also start addressing #2 (still in progress, but we've added more event types already). Thanks for checking it out! We're very open to thoughts / feedback. https://bit.ly/3MGnqvP February 26, 2026 at 12:17AM

Show HN: Tesseract – 3D architecture editor with MCP for AI-assisted design https://bit.ly/46ZURjL

Show HN: Tesseract – 3D architecture editor with MCP for AI-assisted design Hey HN. I'm David, solo dev, 20+ years shipping production systems. I built Tesseract because AI can analyze your codebase, but the results stay buried in text. Architecture is fundamentally visual — you need to see it, navigate it, drill into it. So I built a 3D canvas where AI can show you what it finds. Tesseract is a desktop app today (cloud version coming) with a built-in MCP server. You connect it to Claude Code with one command: claude mcp add tesseract -s user -t http http://localhost:7440/mcp I use it for onboarding (understand a codebase without reading code), mapping (point AI at code, get a 3D diagram), exploring (navigate layers and drill into subsystems), debugging (trace data flows with animated color-coded paths), and generating (design in 3D, generate code back). There's also a Claude Code plugin (tesseract-skills) with slash commands: /arch-codemap maps an entire codebase, /arch-flow traces data paths, /arch-detail drills into subsystems. Works with Claude Code, Cursor, Copilot, Windsurf — any MCP client. Free to use. Sign up to unlock all features for 3 months. It's early but stable. I've been dogfooding it on real projects for weeks and it's ready for other people to try. Demo video (1min47): https://youtu.be/YqqtRv17a3M Docs: https://bit.ly/3OAdPXY Plugin: https://bit.ly/4rCl6VF Discord: https://bit.ly/46qBRL6 Happy to discuss the MCP integration, the design choices, or anything else. Would love feedback. https://bit.ly/4rNDB9G February 26, 2026 at 12:05AM

Show HN: Context Mode – 315 KB of MCP output becomes 5.4 KB in Claude Code https://bit.ly/4sePGF4

Show HN: Context Mode – 315 KB of MCP output becomes 5.4 KB in Claude Code Every MCP tool call dumps raw data into Claude Code's 200K context window. A Playwright snapshot costs 56 KB, 20 GitHub issues cost 59 KB. After 30 minutes, 40% of your context is gone. I built an MCP server that sits between Claude Code and these outputs. It processes them in sandboxes and only returns summaries. 315 KB becomes 5.4 KB. It supports 10 language runtimes, SQLite FTS5 with BM25 ranking for search, and batch execution. Session time before slowdown goes from ~30 min to ~3 hours. MIT licensed, single command install: /plugin marketplace add mksglu/claude-context-mode /plugin install context-mode@claude-context-mode Benchmarks and source: https://bit.ly/3MZWN56 Would love feedback from anyone hitting context limits in Claude Code. https://bit.ly/3MZWN56 February 25, 2026 at 07:23AM