William Trolinger
Manager, Applied AI Research at Heartland Forward
About Me
I lead applied AI research at Heartland Forward, building data pipelines, ML models, and AI agents that inform economic and policy decisions. I architect a Python ETL pipeline that ingests public data across all U.S. states, counties, metros, and micros into BigQuery; design agentic systems that translate natural-language questions into SQL queries and plain-English briefs; and build peer-geography models for community benchmarking using KNN on standardized industry employment shares.
I built the automated data pipeline behind the Most Dynamic Metros and Micropolitans series, ranking 900+ areas annually, and the data pipeline and rankings behind Advancing Regional Innovation Economies (JPMorgan Chase & Nasdaq). I am a named co-author on The Economic Case for Investing in Maternal Health and Disparities in Rural Mental Health. I also run AI training for Arkansas rootEd career advisors and served as a tech mentor for 150+ nonprofit leaders at OpenAI's Nonprofit Jam. Currently pursuing an M.S. in Computer Science at Georgia Tech, specialized in Machine Learning.
Recent Builds
CIR Dashboard & Data Platform
The data platform behind Heartland Forward's Center for Investment Readiness: a React/TypeScript dashboard serving 10 heartland communities with 140+ indicators benchmarked against model-selected peers, on a Python ETL ingesting 35+ public sources into BigQuery with a probe-based scheduler that detects upstream data releases automatically. The Agentic SQL Translator below is this platform's AI layer, currently in internal testing.
Discord ops console for pipelines and dev environment
Two Discord bots — one on a Linux server, one on a Windows desktop — running about twenty-five commands between them. They trigger BigQuery pulls and inspect probe baselines, spawn and manage Claude Code sessions, lifecycle a separate containerized bot (start, stop, logs, rebuild, deploy), run load tests against it, poll NWS for tornado warnings near tracked locations, and can remote-shutdown or reboot the host. Both are auth-gated and route via @-mentions when more than one is listening, so long-running jobs and machine state feel interactive from anywhere.
Personal LLM Wiki
An Obsidian vault that an LLM agent maintains as I drop in articles, papers, and notes. Following Karpathy's pattern: the wiki compounds rather than re-deriving knowledge from raw sources on every query.
Agentic SQL Translator
Internal tool at Heartland Forward that translates natural-language questions about U.S. counties, metros, and micros into BigQuery queries and plain-English briefs for stakeholders.
Research I Have Contributed To
Outdoor Recreation
My Webpages
AI Toolkit
My current favorite is Claude. I initially signed up to use Claude Code, but it's become my go-to for general tasks as well.
For coding, Claude Code has been my primary tool — I prefer using it through the terminal. ChatGPT Teams (through work) gets used for quick summaries and scratch answers when I don't want to fire up a fuller workflow.
OpenRouter is what I use to test new model iterations — it lets me spread API credits across different providers without being locked into one. For local model testing, I use Ollama over LMStudio.
Honorable mentions: SuperWhisper or WisprFlow for transcriptions, ChatGPT Images 2.0 for image generation, and Microsoft Copilot (through work) for its seamless 365 and Outlook integration.