What it is
Private search engine over your own X saves: it describes images, fetches link titles, writes a one-line summary of each post, embeds everything locally, and reranks with Jev.
What it does: Cosine similarity picks 120 candidates; one Jev Score per candidate rates relevance to the query; a separate Jev substance score separates real threads from low-value posts; results are split into strong matches and maybe. Only the model calls leave the machine.
Why it belongs in the showcase: A shortlist-then-Jev-rerank shape on personal data, combined with a tunable composite score.
Caveats: Personal tool; author-reported costs; summaries and image descriptions use another LLM.
Evidence
Public MIT README and code (`tweet_index/jev.py` pins the Jev model and refuses moving `latest` aliases); per-step costs measured by the author on about 6,000 posts; eval scripts included.
Caveats
- Personal tool; capture relies on a userscript that copies responses X's own page already receives
- Author-reported costs (about $0.003 per search, a couple of dollars for a full index); image descriptions and summaries use a separate LLM
- Thin star signal at intake (2★, repo created 2026-10-06)
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.