Research assistant for solar investment

If you want to put solar panels into a new market, someone has to read that country’s regulations first. Across ten countries, several languages and dozens of regions, that’s weeks of reading. Asking a general chatbot doesn’t help much: it mixes up laws from different countries and makes up numbers with total confidence.

I wanted the opposite. Something that would rather say “I don’t know” than guess.

How it works

The markets are Brazil, Mexico, Chile, Colombia, South Africa, Kenya, Nigeria, Malaysia, Indonesia and Vietnam, split into 38 regions and 6 things an investor cares about. Sources only come from government and regulator websites. I ran the collection as batch jobs on Northeastern’s H200 cluster.

Before a fact is allowed into the knowledge base, five agents go through it: one collects, one validates, one writes it up, one looks for alternatives and one audits. A fact needs two independent sources to get in. Llama 3.3 70B does the writing and Llama 3.1 8B does the checking, both on Groq.

When you ask a question, it answers only from what got through. If nothing did, it tells you. If the answer came from a Portuguese, Spanish, Bahasa or Vietnamese source, it says it translated. Each region also gets a 0 to 100 feasibility score, and anything filled in because data was missing is marked as imputed.

Getting it working

ChromaDB kept segfaulting on Windows. Instead of fighting it, I wrote a small in-memory search over hashed embeddings and moved on. The UI started as a Streamlit prototype and got rebuilt with a FastAPI backend once the core worked.

It took 1st place. The code, tests and methodology are on GitHub.