{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/6c174a551046421db5b7eabd73394766\" frameborder=\"0\" width=\"1112\" height=\"834\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":834,"width":1112,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":834,"thumbnail_width":1112,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/6c174a551046421db5b7eabd73394766-e145dd9d4ca42d74.gif","duration":315.966,"title":"DeepSearch Agent for Verified Multi-Source Audits","description":"This Loom demonstrates VPSearch, an Anmol Sen-built web search agent designed for the Serbum FDSE role and term assignment. It runs a seven stage pipeline (planning, searching, fetching, re-ranking, probing, synthesizing, verifying) with live phase labels, audits every claim with verbatim URL quotes, detects conflicts between low confidence sources, and supports clarifying questions and Hindi prompts. The interface shows session history, trace evidence, and exports to Markdown, plus a status page with per provider health metrics like latency, recent errors, and which API keys are throttled. It mentions plain Python and AsyncIO with about 400 lines plus tests, no orchestration framework, Vercel front end and Hugging Places back end, and planned next steps including a judge ensemble, translation prep for Indic queries, and an MCP server interface."}