Dalal Ratings — verified signals from market-commentary video

An ETL and verification pipeline for extracting structured, trustworthy signals from financial market-commentary video. Given a YouTube video, the system pulls and normalizes the transcript, runs it through an LLM pipeline to extract structured calls and claims, and independently runs a computer-vision pass over sampled frames to detect and cluster on-screen signals (tickers, charts, overlays).

Those two independent readings — one language-derived, one vision-derived — are fused and cross-checked against each other. Where they agree, the signal is marked verified; where they disagree, it's routed to a mismatch queue for review rather than silently trusted. The goal is a rating system where the labels are reproducible from source video, not hand-curated.

Stack: Python ETL pipeline, OpenRouter-routed LLM calls, custom computer-vision frame clustering, structured signal storage.

Status: <in active research/prototyping — pipeline runs end-to-end on sample videos; not yet productionized for a live dashboard.>