latentAgent showcase
JobMap
Describe the work you want to do, explore matching job postings, and inspect the evidence behind Jev's relevance judgments.
Jev · Sentence embeddings · PostgreSQL · pgvector
Read technical article
Recorded demonstration. Unmute the player to listen.
What it does
I built JobMap to search real job postings by what someone wants to do. A request such as "I want to build AI agents" produces a list you can explore, with posting details and evidence for the suggested matches. Remote-work filtering helps narrow the results.
The recorded walkthrough follows that search through the current app. It opens an Agent Builder posting, checks its evidence, and uses the collapsible filters. The Compare view puts four search methods side by side, while Usage & costs shows the model spending. The walkthrough has English AI narration.
I also compared the search pipeline with three local embedding models. Jev improved graded relevance on this dataset, while MiniLM found about as many strong matches near the top and returned them faster. That result keeps me from treating a convincing demo as proof of better search.
How it is built
The full article explains the retrieval pipelines, blind judging, and the difference between a related result and a job that supports every requirement. It includes the measured results, costs, and reports you can inspect.
Jev ranked jobs better. MiniLM found about as many strong matches.