clocks go back on Sunday so a lot more people will be finishing shifts and locking up in the dark from next week. good time to check anyone working on their own has a way to raise an alert, and that someone will actually see it
@swarogan bit of both. APIs for the general stuff, but anything with client names, bank details etc either stays local or gets stripped out before it leaves. we built File Croc for that bit
This might just be the way. Might vibe code an open source railway that can spin up the required infra on AWS or Azure or Google depending on the best price at the time.
Open source it.
Who's down for getting involved?
F’it, just put your entire stack on the hyperscalers.
Don’t use Superbase, Resend, Heroku Vercel, Railway, Twilio, Loveable, WordPress, Mailchimp,
Just build it all on AWS, Azure, GCP, and Cloudflare.
F’it, just put your entire stack on the hyperscalers.
Don’t use Superbase, Resend, Heroku Vercel, Railway, Twilio, Loveable, WordPress, Mailchimp,
Just build it all on AWS, Azure, GCP, and Cloudflare.
quick one if you redact PDFs. drawing a black box over a name usually just hides it, the text is still there underneath. select it, copy and paste into notepad, if it comes back it's not redacted
File Croc takes it out properly if you want to give it a go: https://t.co/6uTGvQInfd
quick one for anyone running night security or late shifts on their own. agree a check-in time before the shift starts, and agree what actually happens if it's missed. most places do the first bit and forget the second
@mattwatsonkc fair for most people tbh. where open source makes sense is when you need the CRM to do something specific to your industry that HubSpot just won't do. it's why we built Headquarters, but if HubSpot fits you then stick with it
@innopro_ Ollama's great for this. one thing people miss though, offline doesn't help much if the docs you feed it are full of client names, bank details etc and you end up sharing the outputs later. strip that stuff before it goes in and you're properly covered
@iamkylebalmer worth adding to the setup: Settings > Privacy, turn off model training before you upload any real contracts. and if it's a client's contract take their names and bank details out first, it doesn't need them to review the clauses
@somarkai element-level output is really nice. have you tested it much on scanned forms with handwriting? that's where most parsers fall over, insurance claims, HR forms etc
@LearnWithBrij mostly PDFs, voice notes and photos from site. the photos catch people out, faces and GPS data go straight into the index unless you strip them first. it's half the reason we built File Croc tbh, that step kept getting skipped
@Sumanth_077 the layout coords are the bit I'd use most. chunking by element (table as one chunk, caption with its figure) beats fixed-size chunks every time. we're on the other side of this, PDFs to clean text and chunks with personal stuff stripped out: https://t.co/QSldjs0384
@rubenhassid Good shout on turning training off. I'd add one more: swap the real names, bank details and addresses in the contract for placeholders before you attach it. Claude doesn't need them to spot a dodgy usage rights clause
@jerryjliu0@nvidia sub-1B and top of the board is mad. parsing's only half the job though, the docs people actually want in RAG are full of names, phone numbers, signatures etc. would love to see ParseBench score PII handling one day