fforflo

Hi HN,

I run diofanti.org, a free, open-data platform that monitors government spending and operations in Greece. We transform raw public data into a tableau-like dashboard, providing citizens, journalists, and policymakers with transparency and actionable insights.

For the past five years, a Hetzner-hosted Postgres instance and Metabase have served us well. However, I’ve recently been working on adding an AI/Chat interface to enhance the user experience. I can handle the AI engineering, ML ops, and model pipeline setup myself. Initial prototyping has been promising, but the AI infrastructure costs are now growing—primarily for GPU and AI compute (inference and generation, but no fine-tuning yet).

Diofanti is a non-profit, and while we’re focused on transparency and civic engagement, we’re running into limitations around infrastructure costs. Traditional VC funding routes don’t apply since it's not monetizing, and big cloud providers aren’t a good fit either—most want you to join an accelerator that I don’t have the time to go through. Local (Greek) philanthropic organizations and fellowships have been frustratingly slow, bureaucratic, and not particularly helpful.

I’m looking for any suggestions or funding opportunities to help support our AI infrastructure, whether it’s grants, cloud credits, or connections with organizations that support civic tech or AI for good projects.

If anyone has leads or suggestions (or if you have any experience with similar projects), I’d greatly appreciate hearing from you. Feel free to reach out directly (contact info is on my website: tselai.com).

I’ve been working on some projects with Greek-language data and have encountered some interesting challenges with RAG and LLMs. In an English-speaking universe, it's relatively straightforward to have a decent prototype in a short time. But in other languages, it has proven a bit trickier.

I’m curious to hear from others working with non-English languages—what challenges have you faced? Some areas of interest: - Models that are more open to switch language - Availability and quality of language-specific retrieval corpora - Differences in tokenization and embedding quality - Handling multilingual queries and responses - Any workarounds or best practices you’ve discovered

Would love to hear both success stories and pain points.