Ahhh the old infamous Turbo Tax, commiserations!!
It wasn’t 2 week full effort.
Maybe a day or 2 so spent on another project.
And maybe 2 days “lost” to short days (kids/family stuff).
I maybe had 1-2 hours of collection to do, mostly already done throughout the year.
So I reckon about 6 days?
Definitely not paying off in a year, ROI in maybe year 3 or 4 onwards?
It was complicated due to leaving full time employment, having a study period (Chaincode’s BOSS program) and then contracting (starting up as a sole trader). So 3 different “regimes” to deal with. And one of the main challenges being working out deductions due to running servers where I self host LLMs and dev VMs etc - changing electricity rates and consumption.
I also want to be able to deal with the variety of docs that information needs to be extracted from automatically, but privately. Of course throwing them into something like Claude or ChatGPT is going to extract the information just fine, but I don’t want that leakage of private information.
So that leads to one of the key features being the leveraging of self-hosted Visual Language Model to do the extraction and information population. With prompt tweaking, and at this stage a number of fixed prompts for specific document types, this yields good results (and I’ve learnt just how sensitive the smaller self hosted models are to the prompt). FWIW I’ve been using Qwen3-VL-32B served via ollama for this purpose.
My wife and friends have asked me “can you make money off this”? I reply, probably, but I also absolutely do not want to be developing, maintaining or supporting tax software.
I did this for my own circumstances and because I wanted to get more mileage with coding agent tools and workflows.
Could it have been done quicker, yes.
Has it paid off, not yet, a few years still.
Am I glad I did it, hell yes!!
