← back to brain dumps

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'.

  1. search across llm outputs in web - probably an extension might help
  2. group buy calculator for google play - need to have a splittin' mechanism
  3. cli application for tradin' - cursor for tradin', somethin' which will help traders
  4. link picker - pick all links from a site as an extension (reddit, discord)
  5. awesome list to bookmark - load awesome list to bookmark in chrome and vice versa
  6. openrouter for tools - kinda ipaas
  7. unnecessarily overengineered notetaker
  8. produce a random coin flipper - there is no thing as random for computers, so there will be multiple algos for no reason
  9. exploited tutorial hell - just a bunch of high quality tutorial lists
  10. abusin' technology project
  11. db cleaner - auto sql with gut feelin's
  12. db backup tool - a cli tool which will ask to connect to the db with type and then ask for the output format
  13. kaggle model test script generator
  14. a researcher for finetunin', prunin', distillin', quantizin' - immediately collect and do after releasin' the model automatically
  15. a repo with all possible rag techniques
  16. open notebook lm
  17. monkeytype mcp or ai in there somehow
  18. shellgpt with better ui and fast
  19. a proper implementation of web search in terminal
  20. vim based browser
  21. 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
  22. somethin' better than pydantic - it's just a stupid if condition check
  23. a tool if you pass github url, clone it - a terminal tool maybe
  24. agentify - versionin' for all of the components in an agent
  25. build a fuckin' agent - a simple interactive page to choose the frameworks for buildin' the agent
  26. 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
  27. website to cli tool
  28. a standard benchmarkin' platform with tools for benchmark llm, stt, tts, realtime, embeddin'
  29. bolna.ai clone - with better voice handlin' and meet functionality
  30. ai assist in jupyter notebook
  31. examples finder - find example/cookbook/notebook/recipe folders in github trendin' repos
  32. a collection of pytorch models - model design, blocks, trainer, optimizer, tokenizer, dataset prep, inference
  33. esp32 mcp - connect claude to your home
  34. comparison of tables in vscode (sqlite, csv, excel, parquet) - preview changes
  35. rag engine - a python library with splittin' strategies and similarity metrics
  36. a framework to connect your agent to channels
  37. vibe kanban inside vscode
  38. native voice for openclaw - talk to claude code without typin' like a caveman
  39. 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?"
  40. 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
  41. 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
  42. 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

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.