TruthSeeker

Background

Manually checking claims online means opening ten tabs and reconciling them yourself. The tedious part needed automating: search for evidence, have a model assess it against the claim, and show the sources so I can disagree with the verdict.

Solution

Streamlit web app plus a terminal version. Submitting a statement triggers a Brave Search lookup, and a DeepSeek model reasons over the results to produce a verdict with explanations and direct source links. Search history exports to multiple formats, and Pydantic models keep responses structured.

Process

I kept search and analysis as separate components so I could test each cleanly. The search client has caching and retry logic; the AI can request follow-up searches when initial results aren't sufficient. Most iteration went into result presentation.

Final Product

Reflection

Coordinating external APIs around a model's judgment was the interesting problem, especially letting the model decide when it needs more searches. Filtering source reliability better would be the next investment.