Market Smart — solo AI build, 196 countries

The premise

I wanted direct proof — for clients and for myself — that I can ship a useful AI product end-to-end without a team. Market Smart is that proof: a system that pulls structured event data from GDELT, indexes it for retrieval, and uses the Claude API to produce grounded market summaries across 196 countries in 8 languages.

Stack

• Backend: FastAPI, Python
• LLM: Claude API with prompt caching and tool use
• Retrieval: RAG over GDELT event records
• Frontend: React
• Languages: EN, ES, FR, DE, RU, JA, UK, PT

github.com/accipio09/market-smart

What was hard

• Citation discipline. The summaries had to ground every claim in retrieved events, not in Claude's pretraining. That meant designing the prompt and the retrieval contract together — most of the engineering work was in that interface, not in either piece alone.
• Multilingual without quality drift. Generating in 8 languages from English-heavy retrieval results meant carefully separating the content prompt from the language instruction, and validating that the citations still pointed at the right events.
• Cost. Prompt caching plus a tight retrieval window kept per-summary cost reasonable enough to actually run the thing.

Why this matters for a client

If you're building an AI product and you don't already have a team, the question is whether one person can credibly own the retrieval, the model interface, the data plumbing, and the UX together. This project is the demonstration that I can. Yours will look different — but the shape of the work is the same.