Plain-language tools for understanding policy & power

Leroy's Tools

Most of what shapes our lives is buried in legislation, filings, and reports that are hard to read and easy to ignore. I build open-source tools and automated pipelines that pull out what actually matters and explain it in plain English. Everything ships with the work behind it, so you can check it yourself.

Austin Smith · 19D Army Veteran · National Park College
Started writing Python: October 2025
Available for freelance development, investigative data support, and consulting.

What I've Built

Three open-source projects and one paid pipeline. The tools are free — try Bill Translator in your browser, or clone any project and run it with your own Anthropic API key, no limits.

🔧 Friction Breaker
OPEN SOURCE

Paste any news article, executive order, bill, or legislative text. GLiNER2 extracts entities locally at zero cost, the classifier matches them against a taxonomy of 76 documented accountability-bypassing mechanisms across 8 categories, and Claude generates ranked countermeasures — hardest to reverse first — in plain English. Open source and BYOK: clone it and run locally with your own Anthropic API key.

→ GitHub
📄 Bill Translator
FREE OPEN SOURCE

A legislative transparency tool built specifically for Arkansas Act 602's 8th-grade Flesch-Kincaid readability requirement. Rewrites ballot titles and legislative text to hit the FK threshold while preserving legal force — "shall" stays as "must," section references survive verbatim, and the tool flags any defined terms that drop out between the original and the rewrite. The grade level is scored locally with the same formula the AG's office uses. No key required to try it.

→ Open Bill Translator → GitHub
📡 The Regulated Friction Project
OPEN SOURCE

The research foundation everything else is built on. Tracks structural leverage mechanisms — capital flows, regulatory capture, ballot suppression, energy infrastructure — and documents convergence patterns across nodes. Includes an Oracle single-point-of-failure analysis, an Arkansas infrastructure forensic audit, convergence reports, and verification scripts. Every finding is reproducible. The taxonomy driving Friction Breaker comes from this work.

→ GitHub

Automated OSINT Intelligence Pipeline

Ten automated stages turn three API keys into a live analytical dashboard — no servers, no databases, no DevOps. Three-pass investigation, web-grounded fact checking, and rhetoric-vs-reality gap analysis, all running twice daily on GitHub Actions. Configure it for any intelligence topic in under 15 minutes.

Buy Now — $50 $50 / one-time
01 //
Node Tracker
Queries current status of configurable leverage nodes via Brave Search, then sends results to Claude for structured analysis. Classifies events as FRICTION, COMPLIANCE, ESCALATION, DE_ESCALATION, or NEUTRAL. Identifies leverage type and checks for under-the-radar indicators including regulatory capture and revolving-door personnel moves.
Brave Search + Anthropic Claude
02 //
Dual-Engine Entity Extraction
GLiNER2 runs locally at zero API cost for named entity recognition. Optional Llama Scout 17B cross-validates with structured JSON output — entities, relationships, temporal markers, and cross-node links. Results are merged when both engines are available.
GLiNER2 (local) + Llama Scout
03 //
Convergence Detection
Identifies multi-node convergence patterns from historical run data. Uses adaptive time windows — 1-day for kinetic nodes, 7-day for financial and information nodes. Flags friction↔compliance pairs when attention-capturing and structural events coincide.
Local · zero API cost
04 //
Daily Intelligence Scan
Three-pass investigation protocol: signal sweep across all nodes, cross-node convergence scan for multi-node stories, and under-radar probe for developments not in major headlines. Includes breaking news detection and prediction verification.
Brave Search + Anthropic Claude
05 //
Web-Grounded Fact Checker
Three-phase editorial review: inline date/dollar triage flags future dates and ambiguous attributions, then Brave Search retrieves live web evidence for each claim. Claude verifies, deduplicates, and corrects errors in-place — the dashboard always shows verified content.
Brave Search + Anthropic Claude
06 //
Rhetoric vs. Reality Analysis
Three-column gap analysis: what officials said (rhetoric), what filings and documents show (reality), and what the domestic cost is. Tracks institutional bypass and ideological drivers with confidence levels.
Anthropic Claude
06.5 //
Knowledge Base Context Builder
Ingests entity data, convergence data, and rhetoric/reality output to generate structured knowledge base files — ready for AI agent integration. Auto-generates markdown context for chatbot and assistant workflows.
Local · zero API cost
07 //
Pipeline Validation
Cross-pipeline consistency check — verifies critical nodes appear in top developments, convergence events have entity overlap, and source diversity is maintained. Non-blocking; logs warnings only.
Local · zero API cost
07.5 //
Chart Data Builder
Aggregates full run history into chart-ready JSON for the dashboard's Charts tab — thermostat timeline, stacked area, and node activation heatmap data.
Local · zero API cost
08 //
Interactive Entity Graph
NetworkX builds a directed relationship graph from extracted entities. PyVis renders it as an interactive HTML visualization — color-coded by entity type, embedded directly in the dashboard.
Local · zero API cost
📊
Nine-Tab Live Dashboard
Node Status, Intelligence, Convergence, Predictions, Entities, Entity Graph, Charts, Rhetoric vs. Reality, History. Hosted free on GitHub Pages — no backend required.
🔁
Fully Automated
GitHub Actions runs the full pipeline twice daily at 08:00 and 20:00 UTC. No manual intervention, no servers to manage, no cron jobs to maintain.
🔍
Three-Pass Investigation
Every run executes a signal sweep, cross-node convergence scan, and under-radar probe. Finds stories other tools miss — court dockets, SEC filings, and quiet regulatory moves.
✅
Web-Grounded Fact Checking
Searches the live web for evidence before verifying claims — not just AI knowledge. Deduplicates, corrects errors in-place, and flags ambiguous attributions automatically.
🎯
Prediction Tracker
Logs forecasts and automatically verifies them against live data on every run. Tracks pending, confirmed, and failed predictions with full audit trail on the dashboard.
🛡️
Self-Healing Pipeline
Stages 2–8 are non-blocking — if one fails, the rest continue. The pipeline never fully breaks due to a single stage error. Warnings logged, never fatal.
🧠
Zero-Cost Local AI
GLiNER2 runs entirely on GitHub Actions CPU — no API key, no cost. Dual-engine mode activates automatically when Llama Scout key is provided for cross-validation.
💰
API Budget Controls
Built-in daily call limits with automatic budget files. Prevents runaway API costs — total under $5/month.

What You Need

Three API keys and a free GitHub account. Operating cost under $5/month.

Required
GitHub Account
Free tier is sufficient
Required
Brave Search API
Web search for stages 1, 4, 5
Required
Anthropic API
Claude Sonnet for stages 1, 4, 5, 6
Optional
Llama Scout API
Cross-validates entity extraction
Setup Time
Under 15 minutes
Add keys, push to repo, done
Technical Level
Beginner–Intermediate
Git basics helpful but not required
$50
One-time · lifetime access · no subscriptions
Buy Now
  • Full pipeline source (.tar.xz)
  • Ten-stage automation scripts
  • Three-pass intelligence protocol
  • Web-grounded fact verification
  • Nine-tab interactive dashboard
  • Dual-engine entity extraction
  • Interactive entity graph (PyVis)
  • Prediction tracking & verification
  • AI knowledge base generation
  • Complete setup documentation
  • Email support included

Get in Touch

Questions, ideas, tool issues, consulting inquiries, or just want to reach out — all of the above are welcome. I'm also available for freelance development, investigative data support, and custom pipeline builds.