Selected Work

Coach Artie—a Discord-native AI assistant with persistent memory, tool usage, and multi-modal capabilities. Not a chatbot; a team member.

Key capabilities:

• Persistent memory via PostgreSQL/Supabase
• MCP tool integration for real-world actions
• Multi-channel awareness (Discord, SMS, email)
• Evolving personality through interaction

Technical Architecture

Built on Node.js with Discord.js for real-time interaction. Uses both OpenAI and Anthropic APIs depending on task. Memory system stores conversation context and can recall relevant history. MCP integration allows Coach Artie to take actions—scheduling, file operations, API calls.

Current State

V2 in active development with expanded MCP tooling. The memory system now spans 2+ years of context—enough that Coach Artie remembers project decisions from months ago and can reference them appropriately.

Why This Matters

This is the future of creative work. Not AI replacing humans, but AI that accumulates context and history alongside the team. Coach Artie remembers what we tried, what worked, what failed. He's become institutional memory.

And the research angle: when an AI system has persistent memory over years, what emerges? I'm documenting everything—applying scientific rigor to questions most people wave away.

I documented the complete methodology so other journalists can replicate this analysis for their own police departments. The same techniques work anywhere you have officer complaint data—which, thanks to state transparency laws, is increasingly common.

Open-source accountability tools. That's the point.

I proposed and led development of the Big Board—the touchscreen election visualization system Steve Kornacki uses on MSNBC. Built entirely with web technologies (HTML, CSS, JavaScript, D3), which NBC believes was an industry first for broadcast election graphics.

Key capabilities:

• Multi-touch zoom to view any district, no matter how small
• Historical data for any race in the past decade
• Real-time data integration with election feeds
• Smooth enough for live TV (no lag, no crashes)

Technical Details

The system runs as an HTML5 web app in Chrome kiosk mode on a commodity gaming PC (Intel i7, ATI Radeon GPU), paired with large Planar touchscreen displays. D3.js handles all visualizations and animations.

Same codebase works on-air and embedded on NBC's web properties.

Results

• 19M+ viewers on election night 2018
• Zero crashes across multiple election cycles
• Featured in The New York Times: "A newly juiced-up model of the board that can zoom in on the most obscure House districts"
• Vulture: "Kornacki, eli5-ing races with sleeves rolled up and an 82-inch vertical touchscreen... looked amazing with the board, panning and zooming"

2016 NBC News County-Level Election Maps

Dataproofer—a "spellcheck for data." Drop in a CSV, get an instant report on data quality issues. Open source, runs locally, no data leaves your machine.

Checks for:

• Missing/duplicate column headers
• Empty cells and missing rows
• Outliers from mean and median
• Values at SQL integer limits
• Invalid coordinates
• Character encoding issues

How It Happened

Gerald Rich and I pitched the idea to the Knight Prototype Fund while at Vocativ. We received a $35,000 grant and built it over 6 months with human-centered design training from Knight.

Results

• Open source with community contributions
• Used by newsrooms, researchers, and data analysts
• Part of a broader effort at Vocativ that took graphics output from 5-6/month to 5-6/day

❯ dataproofer data.csv
 
total rows 5035
rows sampled 1259
 
Missing or duplicate column headers: passed
Empty Cells: warn
Duplicate Rows: passed
Outliers from the median: info
Invalid coordinates: passed
 
75%
9 tests passed out of 12
 
### PROOFED ###