SignalWeave

Problem

Keyword search treats every mention as its own hit. That is useful if you already know the name of the thing. It misses the group forming underneath, which is the only moment a weak signal is still early.

Goals

1. 01Ingest RSS feeds and embed the items.
2. 02Cluster related mentions, then score a cluster as growing, steady, or quiet.
3. 03Show the evidence: how recent it is, and how many sources it comes from.

Architecture

SentenceTransformers embed each item. Qdrant stores the vectors and supports semantic search next to keyword match. Gemini writes trend titles and explanations. NetworkX and PyVis draw the relationships. GitHub Actions runs a daily ingest. Streamlit is the dashboard.

Implementation

The pipeline ingests feeds, embeds them, and clusters related signals. A cluster is scored for emergence. The dashboard adds search and filters. Qdrant is the persistent vector memory, so a later run can meet what was already stored. That is the memory described in the Convolve 4.0 record of the project.

Tradeoffs

• Promote a trend only after evidence accumulates

  Recency, source diversity, and coherence have to agree. A single excited mention does not get to become a trend.

• Qdrant as the memory

  Embeddings and cluster centroids live in Qdrant. Without that, every run would meet the feeds as if it had never seen them.

Results

Second place in the Multi-Agent Intelligence Systems track at Convolve 4.0, a global Pan-IIT hackathon, in February 2026, among more than 8,500 teams from around the world. Qdrant and Lords Institute of Engineering & Technology both record the placement. No accuracy figure is published, so none is claimed here.

Limitations

There is no published false-positive rate or coverage number. The detector repository is not public. The landing page source is.