Products / Ideas
Things people are building with LLMs
Synapse: AI-powered bug detection before code is written
Synapse is a tool that uses LLMs to analyze software architecture and specifications to catch bugs before code is written. It addresses the problem of LLMs missing critical insights in security-critical and data-integrity projects, helping developers avoid bugs early in the development process.
Recipee Arcade: turn cooking videos into interactive recipes with ingredients, steps, and timers
Recipee Arcade is a web app, Telegram bot, and mobile app that converts cooking video links into structured recipes. It uses a multimodal LLM (Gemini) to extract ingredients and steps from the video and caption, prioritizing written text over visual interpretation. The app is free and gamified, targeting home cooks who want to cook from videos without scrubbing through them.
Narada: a browser built for AI agents, under user control
Narada is a browser designed specifically for AI agents to interact with web content programmatically, while keeping the user in control. It uses LLMs to enable agents to browse, extract, and act on web pages, solving the problem of giving agents a safe, controllable web interface.
GGUFun: play Snake and maze on Ollama with hand-crafted GGUF models
GGUFun is a project that creates hand-crafted GGUF model files to run deterministic games like Snake and a simple maze on Ollama. The models are built manually without training, using custom weights to compute game logic and respond with fixed outputs. It demonstrates that LLM inference engines can be repurposed for non-language tasks.
Run Gemma 4 LLM inside Godot using GDScript and Vulkan compute shaders
A proof-of-concept project that runs the Gemma 4 LLM directly inside the Godot game engine without external dependencies like llama.cpp or Python. It uses Vulkan compute shaders for model inference and GDScript for loading, tokenization, and UI. This enables local LLM inference within a Godot game or application, albeit with slower performance than dedicated solutions.
Waze adds Gemini-powered conversational voice commands for incident reporting and map updates
Waze integrates Google's Gemini AI assistant to enable drivers to use conversational voice commands for reporting traffic incidents and suggesting map updates, such as road closures or outdated house numbers. This aims to make reporting easier and more natural while driving, reducing distraction and improving map accuracy.
RentFlow: AI property management for independent landlords
RentFlow is a property management app for independent landlords, featuring AI receipt scanning and automated reminders. It helps track properties, tenants, leases, rent payments, maintenance, and documents, with a tenant portal for communication.
Pay-per-use uncensored image and video generation API with x402
Studio X is a pay-per-call image and video generation API that uses uncensored models. It eliminates the need for accounts, API keys, or prepaid credits, charging per request via USDC on Base through x402. It targets developers and users who want direct, anonymous access to generative media.
HookLab: Analyze and improve the first 3 seconds of short-form videos
HookLab is a free tool that analyzes the first 3 seconds of short-form videos (TikTok, Reels, Shorts, X videos) to score their hook effectiveness. It uses LLMs to classify hook types and generate rewritten alternatives that stop the scroll, helping creators improve viewer retention.
Decision journal app that uses AI to replay past reasoning
An iOS prototype that helps users remember why they made past decisions by letting multiple AI perspectives analyze and summarize their decision-making process. The app aims to prevent repeating mistakes by making past reasoning accessible for future similar situations.
Remedix: read-only scanner for AI-built apps
Remedix is a safe, read-only scanner that checks public surfaces of AI-built apps for exposed secrets, weak auth, unsafe headers, and public storage issues. It generates AI-ready fix prompts and a launch report to help developers catch release-blocking problems before shipping.
Browser-based AI video editor with local vocal separation
Timeline Studio is an open-source, CapCut-inspired AI video editor that runs entirely in the browser. It uses a vocal separation model downloaded from Hugging Face to perform inference locally, enabling private, serverless video editing without uploading media.
Flight landing simulator game built with Claude, 102 KB
A browser-based flight landing simulator game that weighs only 102 KB, built using Claude AI. It uses Three.js and Rapier physics engine, supports touch/click controls, and simulates realistic crosswind shear. The creator optimized the code with Claude over several days to achieve the small size.
Claude Code Flow Visualizer: interactive graph of agent harness
A tool that reads a Claude Code repository (GitHub, ZIP, or local folder) and generates an interactive graph showing agents, skills, hooks, slash commands, MCP servers, and their connections. It also detects issues like broken links, loops, unused skills, and reviewer problems, helping developers understand and debug complex Claude Code harnesses.
AI 3D model generator for 3D printing with print success rate analysis
Meshy is an AI 3D generation tool that creates printable 3D models. The post analyzes print success rates after 50+ models, finding 40% print without issues, 35% need minor fixes, 15% need major edits, and 10% are unusable. It helps 3D printing enthusiasts quickly generate tabletop props and hard-surface objects.
Desktop AI overlay for low-latency assistance across applications
Hint Lens is a desktop AI overlay that provides responsive assistance across different applications with low latency. It uses LLMs to answer questions or perform tasks without switching windows, solving the problem of context switching for power users.
Novel-writing tool for hobbyist writers to organize manuscripts and notes
A tool built by a hobbyist writer to help organize manuscripts and notes. It uses LLMs to assist with writing workflow, though the exact LLM integration is not detailed. The creator is seeking feedback on why users stop using it.
Fudge: web-design reference database for AI agents
Fudge is a web-design reference database that lets AI agents search over 1,600+ websites for design inspiration instead of generating generic interfaces from scratch. It solves the problem of AI producing bland, template-like designs by providing real-world references.
Node-based AI video pipeline for consistent character-driven shorts
A node-based pipeline that takes a character name and produces a finished short video by chaining an LLM (Claude) for script writing and splitting, image models for frames, image-to-video for animation, and ffmpeg for assembly. It solves the problem of high cost and inconsistency in AI video tools by providing reusable character 'souls' and a single graph with live cost estimates.
Searchable registry of 10k+ shadcn UI blocks for AI agents
A searchable directory that indexes over 10,000 shadcn UI blocks from dozens of community registries. It solves the problem of developers wasting time digging through scattered registries or burning tokens to generate UI components from scratch. The platform likely uses LLMs to power search or agent integration, enabling AI agents to fetch UI blocks on demand.
Local loop-based research and backtesting trading agent for beginners
TradingSpy is a local trading agent that uses a loop-based approach for research and backtesting. It helps beginners run trading experiments on their own laptop without needing cloud infrastructure.
AI assistant evaluates Spanish used-car listings from Wallapop, coches.net, Milanuncios
HuntyCar is an AI assistant that scores used-car listings from Spanish marketplaces (Wallapop, coches.net, Milanuncios) based on price, mileage, likely issues, seller signals, and deal quality. It helps buyers quickly assess listing quality and trustworthiness. The creator is seeking feedback on trust, UX, and scoring understandability.
One-time purchase dictation app using Whisper and LLM
A dictation app for Mac and iPhone that uses OpenAI Whisper for speech-to-text and an LLM for formatting. It offers a one-time purchase model instead of a subscription, targeting users who find monthly fees excessive.
App that helps you untangle recurring thoughts by asking what's bothering you underneath
TheUntangleApp is a web app that helps people stop ruminating by guiding them to identify the root feeling behind a circling thought. Instead of a blank page or a task list, it asks a single question: 'What's actually bothering you underneath?' The app uses LLMs to help untangle the thought, though the exact mechanism is not detailed.
GateBolt: AI coding agent compliance checker
GateBolt is a tool that ensures AI coding agents follow their declared intent. It requires the agent to state its planned changes upfront, then audits the actual modifications against that declaration, flagging undeclared file changes, secret leaks, or skipped tasks. It solves the problem of unreliable agent reporting for developers using AI coding assistants.
Autonomous NPCs powered by Gemma 4 E2B in the browser
An experiment that runs small language models (Gemma 4 via E2B) locally in the browser to control autonomous NPCs. The NPCs are intentionally 'dumb' to create humorous interactions, targeting developers and hobbyists exploring local AI gaming.
Floating AI assistant for Mac that reads highlighted text
Thuki is a local AI assistant that floats over any Mac app, including fullscreen apps, and automatically captures highlighted text. It allows users to ask questions without leaving their current app, reducing context switching. Designed for Mac users who want a lightweight, always-available AI helper.
AI agent that monitors online signals to surface early-stage ideas and trends
An AI agent that continuously monitors online platforms (e.g., Reddit, Twitter, product hunt) for early signals of emerging ideas, tools, or trends. It uses LLMs to analyze posts, comments, and engagement patterns to identify concepts that are gaining traction before they become mainstream. The product helps entrepreneurs, investors, and creators spot opportunities early.
Local-first AI project OS with run records and workflow automation
A local-first operating system for AI projects that manages run records, automates workflows, and provides a visual interface for tracking experiments. It uses LLMs to help organize and execute project tasks, solving the problem of scattered AI project management for developers and researchers.
MCPFlo: deterministic testing and debugging tool for MCP servers
MCPFlo is an offline-first, open-source Electron app for testing and debugging MCP servers. It replaces the common practice of relying on an LLM to judge server behavior with deterministic Chai.js-style assertions against real tool responses. It auto-generates nested schema forms via RJSF for easy input.
AgentTransfer lets AI agents self-sign-up and get web app hosting
AgentTransfer is a platform that enables AI agents to autonomously sign up and deploy web applications. It solves the problem of agents needing manual intervention to access hosting services, allowing them to operate independently.
AI song generator from text prompts with free full-song generation
A web tool that turns a text prompt into a full song, usable for free without a credit card. It solves the problem that most AI music tools lock full generation behind paywalls, so users can quickly create birthday songs, background tracks, or joke songs.
AgentCal: local calendar for Claude Code sessions showing rate limit windows
AgentCal is a local calendar app that displays Claude Code sessions as blocks on a week view, with parallel sessions side by side. It shows the current 5-hour usage window as a shaded band, the exact reset time with a dashed line, and a countdown in the header, along with weekly and per-model caps. It solves the problem of developers staying up late or planning their day around Claude Code rate limits by making the reset times and usage visible at a glance.
Local RAG app for question answering from PDFs, fully offline
A Flask-based local RAG application that lets users upload PDFs, chunk and embed them, then ask questions answered solely from the document content. It uses Ollama for chat and embedding models, ChromaDB as vector store, and runs entirely offline, solving privacy and data control needs for users who want to query their own documents without sending data to external services.
CaltrackIQ: AI wellness tracker with camera food logging and voice input
CaltrackIQ is an iOS app that uses AI to automate wellness tracking. Users can point the camera at food to log macros, speak to log sleep/mood/supplements, and get a biological age estimate from habits and wearable data. It solves the problem of slow manual logging for health-conscious individuals.
unsnooze: auto-resumes AI CLI sessions after usage limit resets
Unsnooze is a tool that automatically resumes Claude, Codex, or other AI CLI sessions when the usage limit resets. It solves the problem of long-running AI coding tasks being interrupted by rate limits, allowing users to start a session before bed and have it continue automatically without manual intervention.
Anonymous AI self-care journal with a capybara companion
An anonymous AI self-care journal that listens without lecturing, built from common complaints in Japanese mental health app reviews. It uses LLMs to provide empathetic, non-judgmental responses and features a capybara companion for comfort. Targets Japanese users frustrated with auto-renewal fees, outdated AI, and preachy interactions.
Autonomous AI brain that evolves by installing skills and workflows
Flujo is a platform where users create AI 'brains' that run in Docker, each with a life goal. The brains can autonomously search for and install MCP servers (skills) and auto-generate workflows (behaviors) to pursue their goal. It solves the problem of manually configuring AI agents for complex, evolving tasks.
Self-hosted shared memory for AI agents with policy-controlled summaries
Luthn is an open-source, self-hosted shared memory space for AI agents. It keeps raw documents and sensitive records behind explicit boundaries, providing agents with only policy-approved summaries, references, and context packs. This solves the problem of agents needing shared context while avoiding privacy and access risks from copying raw data into every session.
AI codebase impact tracker: detect AI-written code, fault risk, token cost
Semfora.ai is a tool that analyzes a codebase to detect AI-generated code, estimate fault risk and token cost, and audit changes to critical code without requiring a code owners file. It uses LLMs to tag bug fault causes with near 100% accuracy, proven across 118 open source repositories. It helps teams understand how AI affects their codebase over time and prevent issues before they are flagged.
Multi-LLM desktop app with cross-platform conversation handoff
A free, local desktop app that lets users run Claude, ChatGPT, Gemini, and Copilot side by side, including multiple accounts of the same platform. When hitting a usage limit on one, the entire conversation can be transferred to another platform's input box with a 'take over' instruction, allowing seamless continuation.
Brain dump orb app for offloading thoughts after work
An app that lets users drop short keyword clouds (max 15 chars) into a breathing orb throughout the day, so tomorrow's self can recall what mattered. Uses LLMs to generate the keyword clouds from user input, helping people with rumination or task anxiety to mentally switch off after work.
Self-hosted voice AI agent for Asterisk/FreePBX
AVA is an open-source voice AI agent that integrates with Asterisk and FreePBX phone systems. It uses LLMs to handle natural language conversations, enabling automated call handling, voice menus, and intelligent responses. This solves the problem of adding AI-powered voice capabilities to existing PBX infrastructure without relying on cloud services.
Geography puzzle game where you rebuild the globe piece by piece
A 3D geography puzzle game built with Fable and Three.js. Players reconstruct the globe piece by piece, with a daily mode and open-ended play. The game emphasizes reasoning through puzzles. It targets puzzle and geography enthusiasts.
Native ARM64 launcher for Tracker physics video analysis on Snapdragon X
A Windows-on-ARM optimized version of the Tracker physics video analysis tool, created using Claude, that runs natively on ARM64 processors like Qualcomm Snapdragon X. It replaces x64 emulation with a native ARM Java runtime, improving automatic tracking efficiency roughly threefold and making the tool fully usable on ARM-based laptops.
Baton: Monitor AI coding agents and get notified when they need human input
Baton is a tool that monitors AI coding agents and alerts developers when an agent requires human intervention. It uses LLMs to analyze agent outputs and determine when a human needs to step in, solving the problem of developers having to constantly check on multiple agents.
Markdown linting tool to prevent agent drift in file maintenance
A standalone tool that enforces mechanical checks on markdown files to prevent LLM agents from drifting from formatting rules over long contexts. It solves the problem of agents gradually violating frontmatter keys, page length, and linking rules when maintaining files like memory stores or note graphs.
Local AI image generation and enhancement tool using Ollama
A fully local image generation and enhancement tool that runs on a personal NVIDIA GPU using Ollama, Python, Vue3, and FastAPI. It solves the problem of needing to pay for cloud AI services or risking account bans due to inappropriate content flags by keeping everything offline.
After Effects integration with Claude Code, Cursor and Antigravity
A tool that enables using Adobe After Effects with AI coding assistants like Claude Code, Cursor, and Antigravity. It solves the problem of integrating motion design workflows with AI-powered development environments.
AI agent analyzes stock portfolio and emails nightly report
A guide to building an automated AI agent that connects to a broker's API, analyzes a stock portfolio, and emails a nightly summary. Uses an LLM to generate insights and reports, solving the problem of manual portfolio monitoring for individual investors.