Zeitgaist

Zeitgaist started immediately after ChatGPT's first public launch, when mainstream assistants still answered from training data and a fixed knowledge cutoff rather than live retrieval. Decision-makers in finance, marketing, and research needed current public-opinion context across Twitter/X, Reddit, Hacker News, Mastodon, Bluesky, and other platforms. Manual monitoring is time-consuming, critical insights in non-English sources are missed, and traditional search lacks temporal context, source attribution, and precise filters. Ask "why did real estate prices surge recently?" against Chinese sources and you get completely different answers than English - geopolitical drivers, local policies, cultural factors that Western media doesn't cover.

Built a unified backend serving two complementary products - an AI chatbot for conversational queries and an analytics dashboard for trend visualization. In product terms, it was a Perplexity-like answer engine with a deeper social-media index and user-controlled time, language, platform, and location filters. The retrieval system uses dense embedding search followed by cross-encoder reranking, which improved internally evaluated top-10 relevance from 72% to 89% over an embedding-only baseline. Initially tried single-stage retrieval but found accuracy degraded on cross-lingual queries - the two-stage approach improved result quality at acceptable latency (~200ms p95 in the retrieval path).