Judd Gurtman
Ann Arbor, MIjudd@juddgurtman.comjuddgurtman.comlinkedin.com/in/judd-gurtman-208769235github.com/juddg-commits
Information Analysis senior (May 2027) who builds and evaluates LLM agents, and checks what the numbers are actually counting.
Education
University of Michigan, School of Information · B.S. Information, Information AnalysisExpected May 2027
Coursework: Data-Oriented Programming (Python, SQL), Statistics and Data Analysis, Information Ethics; this fall: Data Manipulation, Data Exploration, User Modeling, client capstone.
Online courses completed: Google Data Analytics (Coursera); Analyzing and Visualizing Data with Power BI (DavidsonX, edX).
AI projects, built with Claude Code github.com/juddg-commits/year-of-ai
Coach: full-stack AI fitness app · Python, FastAPI, Claude APIAug to Sep 2026
- Built a FastAPI backend (19 routes), an installable phone front end and a Claude agent that acts through 7 tools: it logs workouts, meals and weigh-ins and writes each day's session from logged lifts.
- Scored the agent with a 10-day simulated user (fake clock, spending cap) and 13 automated checks: 7/13 on the first run, 13/13 in 3 of the last 5. 37 offline tests.
- Cut full-price input tokens 96% with a cache-first prompt; the ten-day test fell from $2.07 (run 1) to $1.35 (run 16).
Research agent: cited research briefs · Python, Claude API, web searchSep 2026
- Answers open questions with a brief: researches the parts in parallel, checks every claim against its source quote, and cites nearly every sentence (45 of 46 on the tuned run; 0 invented sources in any run).
- Counted causes before fixing: in three runs, 76 to 88% of weak claims came from quotes cut off at 150 characters; recovering full sentences cut them from 66% to 54% in a saved replay.
- Cut cost per question from $1.68 to $1.25 and runtime from 6m40s to 2m29s (first vs. tuned run) with a per-stage ledger.
Agent fleet · Python, MCPSep 2026
- Built an MCP server that gives an orchestrator my research agent as a tool, with spending caps and approval before paid runs.
Experience
TrueSource, an OnPoint Group company · Data Analytics Intern, remoteJun to Jul 2026
- Audited 29 KPIs across 4 dashboards (31 tabs), traced 7 inconsistencies to formulas and date filters, and made 8 consolidation recommendations; the team added disclaimers to the live dashboards and flagged one for a rebuild.
- Took my first failure-rate estimate to the process owner and re-cut outcomes into 4 categories (handled correctly, intended routing, caller error, true system error); the real rate was far lower, matching the team's.
- Found why the data was hard to read: many call records had no linked Salesforce work order; classified calls by reason with AI and proposed a first-contact resolution metric.
- Validated a Power BI model against Salesforce: sample-tested 10 records, then compared full populations across 2 tables; after my manager's review, changed my verdict and documented what my first check missed.
- Gave five presentations in six weeks, the last one to company leadership.
Jay's Valet, Aspen · Valet lead, running shifts (part-time)2022 to present
Bosq, Aspen (Michelin-starred restaurant) · Food runner and server supportSummer 2025
Labor for a Neighbor, Aspen · Founder: window washing, yard work and hauling for neighborsSummer 2024
Other projects
- Judd Bot (JavaScript, Cloudflare Workers, Claude API): The voice and text assistant on juddgurtman.com. It answers only from a public fact list, cites sources, and passed 100 of 100 questions in its Oct 4 code-graded eval, including prompt-injection attempts.
- Fuel price volatility (Python, pandas): 26,000+ weekly EIA prices; the West Coast's gap to the U.S. average varies over 2x the Gulf Coast's.
- Movie ratings (Python, APIs, SQLite): TMDb rates action and horror 0.7 to 0.8 points above OMDb; on drama they agree within 0.1.
Leadership
Michigan Youth Sports Initiative: volunteer youth flag football coach, team of about 15, since 2024 · Aspen High School: assistant varsity lacrosse coach, roster of about 30, spring 2026
Skills
- AI:
- Claude API (tool use, structured outputs, prompt caching), MCP, multi-agent orchestration, LLM evaluation, Claude Code
- Code and data:
- Python (pandas, NumPy, matplotlib), SQL, JavaScript, TypeScript, FastAPI, REST APIs, Git, GitHub Actions, Cloudflare Workers, Power BI (model validation), Salesforce, Excel
- Interests:
- Skiing, sports analytics, lacrosse, house music