Every morning, before the market opens, Finlio tells you in plain English why your stocks & mutual funds are moving.
No jargon. No broker login. Just one message that actually makes sense.
🎉 Waitlist is open → https://t.co/dWZItm6XNU
been using @paper for designing and framing the ux of my upcoming mobile app but then i feel like the mcp limit is really harsh and us having to wait for 5 days for it to reset seems a bit off, I feel like that alone will make a lot of people not want to continue using the app much. but overall its a solid tool i have been using recently
This is a hero section i designed for my client, I am sure people think claude is bad at designing and yes it is not the best but when there is a human in the loop who knows how to design, everything seems to just fall in place.
Sent Claude Code a screenshot asking it to clean up some UI.
It spotted the "approaching weekly usage limit" banner in the corner, which I hadn't mentioned, and started batching its work to burn fewer tokens.
Didn't ask it to. That's the kind of attention that's hard to fake.
One rough edge from testing: the model emits layout coordinates inside its transcriptions. So Haypile now strips grounding markup before anything reaches the index.
Found the edge and shipped the fix the same evening. Building in public includes the sausage.
I pointed Haypile at Unlimited-OCR, Baidu's hot new OCR model.
Haypile ships no OCR model on purpose. Scanned pages go to whatever vision model your local server runs. It spotted "ocr" in the name and picked it automatically.
Indexed, searchable, page cited. Zero cloud.
ollama pull frob/unlimited-ocr:f16
hay add ~/scans
That's the whole setup. The test file was an image-only scan of a 1987 memo, no text layer. Search by meaning finds it, exact IDs still match, every result cites file and page.
Free and open source (AGPL). macOS, Linux, and Windows, today:
Try it: https://t.co/50rrcdapk9
Code: https://t.co/qn2aWsJEn5
I'll be posting progress, real numbers, and the engineering behind it here. Follow along if local-first AI is your thing.
Your files are basically unsearchable beyond filenames and exact keywords.
Cloud AI fixes that by making you upload everything. Self-hosted RAG fixes it with Python, Docker, and a vector database to babysit.
So I built the missing option: one binary, point it at folders, done. Meet Haypile.
The whole flow. Point it at a folder and it's searchable in seconds.
"agreement cancellation" finds the termination clause even though no file contains the word "cancellation". Search works by meaning, exact identifiers still match exactly.
And hay status reports outbound connections: 0. That's a feature, not a promise.