dump software ideas
ideas i had at 3am or in the shower. most of 'em are never gonna see a git init. but writin' 'em down makes me feel productive without actually doin' anythin'. [✓] = shockingly, i built it. public ones get a link, private ones just get a name, 'cause they ain't ready to be seen by strangers.
- search across llm outputs in web - probably an extension might help
- group buy calculator for google play - need to have a splittin' mechanism
- cli application for tradin' - cursor for tradin', somethin' which will help traders
- link picker - pick all links from a site as an extension (reddit, discord)
- awesome list to bookmark - load awesome list to bookmark in chrome and vice versa
- openrouter for tools - kinda ipaas
- unnecessarily overengineered notetaker
- produce a random coin flipper - there is no thing as random for computers, so there will be multiple algos for no reason
- exploited tutorial hell - just a bunch of high quality tutorial lists ~ [✓] sorta, that's Awesome-GenAI. genai only, 'cause one hell at a time
- abusin' technology project
- db cleaner - auto sql with gut feelin's
- db backup tool - a cli tool which will ask to connect to the db with type and then ask for the output format
- kaggle model test script generator
- a researcher for finetunin', prunin', distillin', quantizin' - immediately collect and do after releasin' the model automatically
- a repo with all possible rag techniques
- open notebook lm
- monkeytype mcp or ai in there somehow
- shellgpt with better ui and fast
- a proper implementation of web search in terminal - [✓] built it as
search-tool. my own searxng, 14 engines, one ranked list, zero api keys. google can track someone else now - vim based browser
-
live experiments - have a stupid question, find a way to experiment it with docker
- for a question - clearly define: question, possibilities, a docker file which can run, 5 runs with logs, verdict
- templates: docker, sh file, makefile, each programmin' language, multiple language, each os, api experiments, llm experiment, mcp experiment
- somethin' better than pydantic - it's just a stupid if condition check
- a tool if you pass github url, clone it - a terminal tool maybe
- agentify - versionin' for all of the components in an agent
- build a fuckin' agent - a simple interactive page to choose the frameworks for buildin' the agent
-
how tech love each other - all the integration guides as a graph viz
- every platform's integration pages mapped as nodes and edges
- e.g. langchain → pinecone edge links straight to langchain's pinecone integration doc
- finally see who's friends with who without diggin' through 12 docs sites
- website to cli tool
- a standard benchmarkin' platform with tools for benchmark llm, stt, tts, realtime, embeddin'
- bolna.ai clone - with better voice handlin' and meet functionality ~ [✓] the voice half is
voice-ai, self-hosted on livekit, callable from a browser or a real phone. the meet half is still livin' in the shower - ai assist in jupyter notebook
- examples finder - find example/cookbook/notebook/recipe folders in github trendin' repos
- a collection of pytorch models - model design, blocks, trainer, optimizer, tokenizer, dataset prep, inference
- esp32 mcp - connect claude to your home
- comparison of tables in vscode (sqlite, csv, excel, parquet) - preview changes
- rag engine - a python library with splittin' strategies and similarity metrics
- a framework to connect your agent to channels - [✓] that's
channels. whatsapp, telegram, slack, twilio, email, the lot. webhooks in, replies out, my sanity sideways - vibe kanban inside vscode
- native voice for openclaw - talk to claude code without typin' like a caveman
-
what i did today - realtime data collector for your dev work
- hooks into github, slack, and the rest of the dev tool circus
- search, summarize, and spit out a dsr without typin' a word
- integrates with meetings so hr finally stops askin' "what did you do today?"
-
open interview notetaker - kinda open source fireflies.ai
- a bot that joins your google meet and zoom calls
- transcribes the whole thing in realtime
- open source so you ain't gotta sell your soul to fathom or otter
-
a framework with native platform agent - every backend/frontend/fullstack framework out there and not one ships an agent by default
- platform agent = an agent that lives inside your app and actually knows your platform - so users can just chat with it to get stuff done instead of clickin' around
- a buncha frameworks are slowly switchin' into this, baking the agent right into the platform instead of bolting on a chatbot later
- not just backend api stuff - it's about every element you build. an api, a frontend button, a form, a toggle, anything a user touches
- when you build the element you just describe what it does, and the agent learns it - "this button exports the report", "this endpoint creates an order", "this form updates the profile"
- so the agent can actually drive the whole app - click the buttons, hit the endpoints, fill the forms - on behalf of the user
- flip the config to true and a chat tab shows up in your fullstack app, no extra wirin'
- basically the openapi boolean vibe - one flag and it's just there
- don't even gotta build a whole new framework - could just be a plugin you bolt onto an existing one (django, next, laravel, whatever) and it brings the agent along
-
pydantic but parallel - validate every field and condition at once instead of one by one
- pydantic bails on the first fail, i want the full list of what's broken
- instead of "x api key missin' from env", show "x, y, z env vars missin'"
- super useful for llm output validation - one round trip to fix everythin' instead of playin' whack-a-mole
-
hackathon farmin' with codin' agents - find every hackathon worth enterin', then let claude code actually build the thing
- step one is boring: an aggregator over devpost, unstop, mlh, hackerearth - filtered by stack, prize, and how much time's left
- step two is the fun part - hand the problem statement to an agent and let it run end to end while i sleep
- open question whether an autonomous submission is impressive or just spam. probably depends entirely on whether it wins
-
cli auth that follows you between machines - log into gh on one mac, sit down at another, already logged in
- every cli hides its creds somewhere different - gh in the keychain, aws in ~/.aws, docker in config.json, npm in .npmrc. no standard, no sync, no mercy
- settin' up a new laptop means re-authin' eight tools through eight browser tabs, every single time
- basically 1password for the tools that never got the memo about oauth
- whole thing lives or dies on the encryption tho. sync the tokens badly and you haven't built a product, you've built a very convenient breach
-
vibe github - vibe kanban but the board is just github, and the tickets run themselves
- github projects is already a kanban board, so quit buildin' another one. just read that one
- anythin' that shows up - a fresh issue, a pr comment sayin' "can you rename this" - gets picked up and actually finished
- the session runs in github actions, so nothin' lives on my laptop. i can shut the lid and the work keeps happenin'
- sessions are pullable tho - if the agent goes off the rails you yank it down and finish it local like a normal person
- end goal is openin' github monday mornin' and half the column already moved to done without me touchin' a key
-
os wide mcp - let the agent see my live chats and every terminal that's actually runnin' ~ [✓] half of it
- every mcp out there reads files and hits apis. not one of 'em can see the log scrollin' in tmux pane 3 right this second
- read the live terminals - tmux panes, iterm tabs, the vscode one i forgot i opened - so "why did it crash" don't need me copy pastin' a stack trace like it's 2011
- same deal for chats. slack, discord, whatsapp, imessage ~ the context is always sittin' in a dm and never in the repo
- the machine already knows what i'm doin'. the agent is the only one in the room that don't
- privacy wise this is a keylogger with good branding, so it stays local or it don't get built
- [✓] the terminal half got built, twice, 'cause one wasn't enough.
agent-eyelets an agent read and nudge the other claude and codex sessions on my mac,herdr-mcplets it boss around every pane in herdr. the chats half is still a keylogger waitin' for a lawyer
- claude drives my actual chrome - my profiles, my logins, and a cursor that moves like a human on mild caffeine ~ [✓] built it as
browser-hand. the bank profile is read-only, 'cause i know myself - a companion app for my car - the venue knows its fuel level and refuses to share ~ [✓] that's
car-mate. logs trips, remembers i parked in b2 again, does km/l so i stop doin' math on receipts - remove claude as co-author from my git history - it wrote the code, i'm takin' the credit. that's just management ~ [✓] claim-your-code. on brew and scoop, so the theft is cross platform
- llm as the teacher, a tiny fast model as the student - fine-tune a decision model on your own data in one line ~ [✓] decisionsmith. on pypi and everythin', like a real adult
- let gemini stare at my dataframe and tell me what's interestin' - skip the part where i stare at it myself ~ [✓] GemInsights. sorry, data analysts
- agents that job hunt for me - find it, research the company, tailor the resume, fill the form ~ [✓]
job-hunter. does everythin' except actually get me hired. i still press submit, like an adult - legal advice from an llm - rag over the legal stuff so it cites instead of vibes ~ [✓] Legal-Advisor-LLM. not a lawyer. neither is the bot. please don't sue either of us
- give the mac notch a job - babysit my ai agents from the top of the screen ~ [✓]
mac-notch. the black rectangle finally pays rent - an expense tracker that finally shows where the salary goes - spoiler: it goes ~ [✓]
money-logs. android, web, and a rust backend nobody asked for - a penpot mcp that don't need a browser tab babysat forever - just talk to the api ~ [✓] penpot-mcp. the official one can keep its tab
- one cli for the whole pypi ritual - uv, poetry, hatch, twine, keyring, all of it, minus the 2am regret ~ [✓] kinda,
pypictl.infoandnamework. the other eleven plans are vibes - an evaluation library for llm and rag apps - is it right, is it grounded, is it rude ~ [✓] ragrank. was perfect when i built it. now it has issues. classic
- click-ops ui for streamin' pipelines - cdc in, sql generated, sinks out ~ [✓]
realtime-datapipes. risingwave does the work, streamlit takes the credit - scripts that click websites so i don't have to - one per website that annoyed me once too often ~ [✓]
scraper-buddies. it's a monorepo of grudges
if you're actually crazy enough to build any of these, ping me. i'll jump in as a collaborator, hype man, or at the very least the guy who opens issues and never closes 'em. seriously tho, don't let these die in my shower thoughts - hit me up.