Cassandra — Stage One Cloud

AI market intelligence pipeline · local inference · no cloud dependency

Active Build Python VADER Ollama Local LLMs NLP yfinance

Origin

Seven months ago I made a small investment based on a pipeline I had built. It is down about 30%. I am keeping it — not as a hope-it-recovers play, but as a reminder that the tool was doing its job. I just had not built the other half yet.

That is the honest version of what led to Cassandra.

The original inspiration was the changing landscape around cannabis and federal banking reform — a real signal worth tracking. But that framing was too narrow. It was limiting the story of what this tool actually is.

Cassandra is not a stock screener. It never was. It is a domain-agnostic intelligence pipeline that combines VADER sentiment analysis with local LLM inference to surface signal from unstructured text — at scale, on your own hardware, with no data leaving the machine.

The name felt right. The prophet no one believed. Dark humor appropriate.

−30%

The investment that started all of this. A small position taken based on an early pipeline run. Still open. Still a reminder.

How It Works

The Pipeline

RSS Collection

Topic-specific and general RSS feeds are scraped across configurable keyword sets. Article bodies are extracted and deduplicated. Output is a timestamped CSV ready for scoring.

VADER Sentiment Scoring

Every article is scored with VADER — fast, consistent, no model required. Compound scores, label breakdowns, and per-topic aggregates give the LLM something structured to reason about instead of raw noise.

Local LLM Analysis via Ollama

Pre-analyzed data is handed off to a locally running language model through the Ollama HTTP API. No external API calls. No data leaving the machine. The model produces a structured narrative intelligence report.

Structured Report Output

A timestamped markdown report is saved covering overall sentiment, narrative summary, stakeholder outlook, and key risks. Domain configurable — point it anywhere text carries signal.

Latest Report

⚾ MLB Lockout — June 2026

MLB Lockout Sentiment Report

Generated 2026-06-07 · Model: qwen2.5:7b · Local inference via Ollama

Full Report →

Articles Analyzed

Overall Sentiment

Negative

Avg Compound Score

−0.1403

Positive / Neutral / Neg

11 / 39 / 37

CBA Negotiations

−0.0978

Revenue Sharing / Salary Cap

−0.0886

MLBPA Players Union

−0.1057

Rob Manfred — Commissioner

−0.4275

Roadmap

Where This Is Going

Phase 1 — Complete

Core Pipeline

RSS collection, VADER scoring, Ollama LLM integration, structured markdown report output. Domain configurable via config block.

Phase 2 — In Progress

True Domain Agnosticism

Moving topic, feeds, and prompt framing from static config to fully dynamic inputs. Point it anywhere, get a structured report back — no code edits required.

Phase 3 — Planned

Social Sentiment Layer

Scraping X, Reddit, and Facebook to pull real-time social sentiment alongside the RSS pipeline. News articles tell one story. What people are saying tells another.

Phase 4 — Planned

Frontend Interface

A web UI so Cassandra does not live entirely in a terminal. Configure topics, trigger runs, and browse reports without touching the command line.

Phase 5 — Considering

External LLM API Integration

Optionally routing to a more powerful model via API for deeper reasoning. Local inference runs first — this gets earned by expanding the data layer, not skipping ahead.

Tech Stack

Built With

Sentiment Analysis

VADER (vaderSentiment)

Compound scoring

Per-topic aggregation

LLM Inference

Ollama (local)

qwen2.5:7b (current)

HTTP API — no cloud

Data Collection

feedparser (RSS)

BeautifulSoup (body extraction)

pandas (CSV pipeline)

Infrastructure

Mercury (Windows GPU host)

Atlas (Ubuntu server)

Python 3.x