Rendered at 13:47:03 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
desktopentree 2 days ago [-]
For the technical audience, the tool runs on Cloudflare Worker with D1 for data and R2 for photos. It's one hand-rolled HTML file and is only ~1.1MB. I built in an HEIC decoder utilizing libheif in a Web Worker. It also has an android and apple version with watch, widgets, and Live Activity native capabilities.
Sonnet 4.5 does the extraction with numerous logic variations built in from trial and error, on many different file types. I am challenging all to try and stump it (PDF, .eml, .csv, .tsv, JPG, .ics).
1 days ago [-]
bodhiriggins 15 hours ago [-]
Isn't this just an LLM call? What's the moat?
desktopentree 14 hours ago [-]
Very fair callout and question.
The extraction setup was actually the easy part. The magic in the tool is from the deterministic rules I put in place over months of iteration using it myself.
Inconsistent date formats, multi-lingual reservations, logic around when a new trip starts and your previous trip ends, parsing both an expense and a confirmation from the same email, duplication rules for when someone accidentally adds the same booking but from a different source, etc. These were all causes of actual parsing errors found from people testing it out.
I built 68 modules within the tool itself from testing.
The logic of the output will only improve over time with more usage. Building the right logic and ruleset into every conceivable scenario is the end goal.
Founderarcstone 2 days ago [-]
Looks clean great work
desktopentree 2 days ago [-]
Thank you! I had a lot of fun branding it around the bird. Do you have any unique booking confirmations to test the tool with?
Founderarcstone 2 days ago [-]
I don't but this was interesting to see as I have not seen this put together like this. I just appreciate the creativity.
Sonnet 4.5 does the extraction with numerous logic variations built in from trial and error, on many different file types. I am challenging all to try and stump it (PDF, .eml, .csv, .tsv, JPG, .ics).
The extraction setup was actually the easy part. The magic in the tool is from the deterministic rules I put in place over months of iteration using it myself.
Inconsistent date formats, multi-lingual reservations, logic around when a new trip starts and your previous trip ends, parsing both an expense and a confirmation from the same email, duplication rules for when someone accidentally adds the same booking but from a different source, etc. These were all causes of actual parsing errors found from people testing it out.
I built 68 modules within the tool itself from testing.
The logic of the output will only improve over time with more usage. Building the right logic and ruleset into every conceivable scenario is the end goal.