I built FinSight because I wanted to analyze my spends, inflow and outflow. But I did not want to upload a statement to a cloud LLM for data privacy.
Finsight provides LLM-assisted transaction categorization without uploading bank or credit card statements to a 3rd Party service.
Architecture: PDF parsing client-side via pdfjs-dist, AI inference via local Ollama/LM Studio API, storage in localStorage/sessionStorage via Zustand. No backend (yet)
A few things I found technically interesting:
1. Context window management is the main challenge with long statements.
I'm handling it by chunking transactions and doing a second pass
aggregation. It works but it's the messiest part of the codebase —
would genuinely value feedback on better approaches.
2. 1B parameter models are sufficient for parsing. 7B models give
meaningfully better categorization accuracy. The main constraint
isn't model capability — it's context window length with large statements and speed.
3. Personally, Qwen 3 gave me the best results but was the slowest in processing a large file. Gpt-oss-20b was faster but the categorization wasn’t as good. Speed is of course, hardware dependent.
3. PDF statement formats vary enormously between banks. LLM-based extraction handles this variation better than any regex approach I’ve tried.
Caveats: setup requires Ollama or LM Studio plus a model download,
which is 20-30 minutes on a fresh machine.
I built FinSight because I wanted to analyze my spends, inflow and outflow. But I did not want to upload a statement to a cloud LLM for data privacy.
Finsight provides LLM-assisted transaction categorization without uploading bank or credit card statements to a 3rd Party service.
Architecture: PDF parsing client-side via pdfjs-dist, AI inference via local Ollama/LM Studio API, storage in localStorage/sessionStorage via Zustand. No backend (yet)
A few things I found technically interesting:
1. Context window management is the main challenge with long statements. I'm handling it by chunking transactions and doing a second pass aggregation. It works but it's the messiest part of the codebase — would genuinely value feedback on better approaches.
2. 1B parameter models are sufficient for parsing. 7B models give meaningfully better categorization accuracy. The main constraint isn't model capability — it's context window length with large statements and speed. 3. Personally, Qwen 3 gave me the best results but was the slowest in processing a large file. Gpt-oss-20b was faster but the categorization wasn’t as good. Speed is of course, hardware dependent.
3. PDF statement formats vary enormously between banks. LLM-based extraction handles this variation better than any regex approach I’ve tried.
Caveats: setup requires Ollama or LM Studio plus a model download, which is 20-30 minutes on a fresh machine.
Installation & Demo Video - https://youtu.be/VGUWBQ5t5dc
GitHub - https://github.com/AJ/FinSight?utm_source=hackernews&utm_med...
MIT licensed.