Flash News Demo (NDA-safe excerpt)

This repository is a small, public demo extracted from a larger private trading / data-quality system (the full system is under NDA). The goal of this repo is to demonstrate Python engineering quality, data validation, and an LLM labeling/evaluation loop in a way a recruiter can run locally in minutes.

What it does

• Fetches crypto-related articles from NewsAPI for a set of tokens.
• (Optional) Labels each article using an LLM into a small taxonomy:
  • historical: True/False (is the article mostly about past events?)
  • sentiment: one of significant rise | rise | neutral | fall | significant fall
• Writes inspectable artifacts (CSV + JSON) under data/news/<TOKEN>/.

What this demo intentionally does not include

• Private strategy logic, proprietary datasets, execution/routing code, infra, monitoring, or any production keys.
• Anything that would reveal client data or confidential decision rules.

---

Quickstart (run in <10 minutes)

1) Install

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

pip install -r requirements.txt
pip install -r requirements-dev.txt  # optional: lint/type-check/tests

2) Configure environment

Copy the example env file and insert your keys:

cp .env.example .env

Required:

• NEWS_API_KEY (NewsAPI)

Optional (only if you run labeling):

• OPENAI_API_KEY
• OPENAI_MODEL (default: gpt-4o-mini)

3) Run

Fetch only (no LLM calls):

python get_news.py --tokens BTC,ETH --start 2024-09-27 --end 2024-10-06

Fetch + label (uses OPENAI_API_KEY):

python get_news.py --tokens BTC,ETH --start 2024-09-27 --end 2024-10-06 --label

Outputs (example):

data/
  news/
    BTC/
      2024-09-27_to_2024-10-06_everything_BTC.csv
      2024-09-27_to_2024-10-06_sentiment_BTC.csv
      2024-09-27_to_2024-10-06_rise_dates_BTC.json
      2024-09-27_to_2024-10-06_fall_dates_BTC.json

Caching behavior:

• If the *_everything_<TOKEN>.csv exists, the script reuses it and does not hit NewsAPI again.
• If the *_sentiment_<TOKEN>.csv exists, the script reuses it and does not call the LLM again.

---

Repo contents

• get_news.py — CLI + core logic (NewsAPI fetch, optional LLM labeling, artifact outputs).
• flash_news.ipynb / flash_news.html / Flash News Presentation.pdf — supporting demo materials.
• tests/ — pytest unit tests for parsing, caching, and API boundaries.
• .github/workflows/ci.yml — CI pipeline (ruff + mypy + pytest) for Python 3.10–3.12.

---

How this aligns to target roles (keywords)

LLM Trainer / Reasoning Specialist

• Constrained taxonomy labeling and robust parsing of model output.
• “Prompt → structured label → artifact” loop suitable for evaluation datasets.

AI Data Evaluator / Data Quality Reviewer

• Clear label schema, deterministic outputs, cache-first runs.
• Tests covering messy model output, empty fields, and boundary behavior.

Technical Writer – AI Training

• End-to-end reproducible README, explicit limitations, and artifact description.

Python Coding Specialist / Debugging

• Separation of concerns (NewsApiClient, labeler interface, analyzer orchestration).
• CI: linting, formatting, typing, tests.

Scientific Coding – Maths & Python

• Dataframe-driven outputs, reproducible runs, predictable artifact naming for downstream backtests.

Tool-Calling / Agent Evaluation

• Clean “adapter” interface (LlmClient) that mirrors tool calling boundaries.

---

Security & NDA-safe notes

• This repo contains no secrets. Add keys only via .env / environment variables.
• .env.example is intentionally non-sensitive.
• The private system includes additional components (data sources, infra, strategy logic) that are not required to understand this demo.

---

Development

Run quality checks:

ruff check .
ruff format .
mypy .
pytest

---

License

MIT License (see LICENSE).