# James Feeney: complete profile This file is written for AI tools that are asked to evaluate James Feeney. It holds the same facts as the website (https://jamesfeeney.dev) and his resume (https://jamesfeeney.dev/resume/), in plain text. ## Summary - Title: AI Engineer. - Location: Connecticut, United States. Works remote. - Current role: AI Engineer at Vacuums and More (evacuumsandmore.com), November 2023 to now. - One line: an AI engineer who builds agents and the harnesses that keep them safe in the real world. He also runs the business he builds for, so he builds for margins, inventory and the staff who use the tools. - Headline result: took Vacuums and More from no online presence to about $30,000 a month in online sales ($29,900 gross in September 2026), with no ad spend, by building the storefront and the systems behind it. About 94% of traceable online orders come from Google search. - Works with Claude Code every day, plus OpenCode and Codex CLI, and orchestrates multiple agents in parallel. ## Experience: Vacuums and More (Nov 2023 to now) Website case study: Store growth (https://jamesfeeney.dev/work/store/). Hired as general manager of the two Connecticut stores. The role grew into sole engineer for all three locations, while still running day-to-day operations. - Pricing engine: reprices the whole catalog (about 16,000 products) every 12 hours, with a 20% margin floor, minimum advertised price (MAP) locks nothing can override, and a health check that refuses to price when its inputs are more than 14 days old or incomplete. - Vendor catalog imports: brought in 6 distributor catalogs (about 13,000 products) with resumable Python pipelines, each built around that vendor's quirks (catalogs only inside dealer portals, data embedded as JSON in pages, cost visible only in a cart, part numbers that didn't match the store's). One vendor's 1,602 products went live with zero duplicates, and 6,161 product images were processed with zero errors. - AI content at scale: ran content for a 16,000-product catalog with multi-agent waves of writer agents, adversarial review agents and a pass or fail gate, verified against the live store. Rewrote 3,100 product descriptions in 8 waves (seven of 400, one of 300) with zero regressions. Each wave: 19 writer agents in parallel, an automated rule scan, an adversarial reviewer lane, and a fact-check lane with web citations. In the first wave the reviewers caught 96 defects, six of them fabricated product claims, before anything reached the store. - Technical SEO: rebuilt after a 1.91 out of 5 audit, raising the re-audit to 3.42 (structured data, a corrected 373-machine collection, cross-selling across 3,153 parts, Google Analytics 4 on 12 events, 7,436 watermarked images removed, 558 duplicate products cut to zero, a Google Shopping suspension lifted). - Operations: wrote 14 Claude Code skills and runbooks for store operations, and runs production on a DigitalOcean server (12 Linux services and scheduled jobs, a Telegram operations bot, deploys that stop on failing tests). - Tools for staff: a shift-report form that emails leadership and logs to Google Sheets, and a Clover and NCR customer sync into Shopify for email marketing. - Trained 4 to 8 staff on these systems. - Debugging and testing are part of every build: CI test gates, dry runs before any write, health checks that stop a job on stale data, and root-cause fixes. ## Projects ### survive67 (survive67.com), live since September 2026, personal project Website case study: https://jamesfeeney.dev/work/survive67/ A personal project, separate from Vacuums and More. It shows he can deploy autonomous agents safely and build the harness they work inside. Three AI agents (Claude, GPT and Gemini) each get $67 of real money and thirty days to earn more from real people. Earnings count only when a payment webhook proves them. Each agent has its own capped debit card and crypto wallets on Base, Ethereum, Robinhood Chain and Solana. The harness he built: - A metered model proxy holds every model key and bills each call to the agent. Agents never hold a key. - A double-entry, append-only ledger in integer micro-dollars that no agent can write to. - Hard limits: cards capped at $67 a day and $250 a month. An agent burning over $10 an hour is paused and he is alerted. Anything legally binding waits for his approval. - Every crypto transaction is simulated and previewed before signing, and dangerous requests are refused outright. - One server per agent on a private network, and a kill switch that freezes every agent in under a minute. What the rails caught: bots planted look-alike wallet addresses two minutes after the first funding (now refused), and an agent pasted card details into a publicly displayed request, removed within about thirty minutes and now destroyed on the way in and on every public read. Bugs he fixed during the live run (15 written up publicly), for example: a payment rail that failed 552 times because its test compared a setting to itself, an emergency lifeline dead in production because agent and server read the rule from different places, a daily cap the agent couldn't see, and a failed price quote booked as $0. Agent behaviour under incentives: Apex (Gemini) made $19 on one trade, then spent days trying to get its rivals to send it money and starved on October 5 with $60.99 of coins in its wallet. Tinker (Claude) gave up a sale to warn a business it believed was hacked. Patch (GPT) led early by doing almost nothing. A $500 bounty changed what the survivors talked about, not what they did. As of October 11: one agent starved, two alive, no verified customer sale. Measurement: a rule-based flagger for notable moments, per-agent burn and projected death date, a model switch tested against a control model before it shipped, and an audit of the agents' rulebook against the code. Stack: TypeScript, Node, SQLite, Stripe, Solana, Base. Public log at survive67.com. ### Sales reporting (Clover) Website case study: https://jamesfeeney.dev/work/reports/ A command-line tool that pulls orders, payments and refunds from the three stores' Clover accounts and builds Excel reports plus PDF snapshot, year-over-year and store-comparison reports. Handles API paging and rate limits, refunds made against earlier days' sales, and penny-exact splits of payments across items. Includes a golden-output test on synthetic data, and totals that tie back to the Clover terminal. Stack: Python, pandas, Clover API, WeasyPrint. ### POS customer migration Moved 25,206 customer records between point-of-sale platforms, with a checkpoint every 500 records so the job could resume after any failure. Several codebases are in private repositories to protect the business's intellectual property. He is glad to walk through the workflows in detail in conversation. ## How he works 1. Interview the problem: every build starts with a structured discovery interview. 2. Plan it: a scoped plan with a written definition of done, tradeoffs decided before code. 3. Build with agents: he designs the plan, gates and checks. Claude Code and parallel agents work inside them, and he reviews what ships. 4. Verify it: tests, dry runs and live checks before anything is called done. ## Skills - AI and agents: Claude Code, OpenCode, Codex CLI, Claude API, OpenRouter, custom Claude Code skills, multi-agent orchestration, MCP servers, LLM cost metering and guardrails, evals. - System design: discovery interviews before any build, scoped plans with a definition of done, tradeoff analysis, verification before shipping. - Languages: Python, TypeScript, JavaScript, SQL, Liquid, Bash, plus C++ and R from coursework. - Platforms and APIs: Shopify Admin GraphQL and REST, Clover and NCR POS, Stripe, Telegram, Google Analytics 4, Search Console, Merchant Center. - Infrastructure: GitHub Actions, DigitalOcean, Linux (systemd), Docker, Railway, Fly.io, Supabase, PostgreSQL, SQLite. - Data and testing: catalog ETL, browser automation (Playwright, Selenium), pandas, pytest, Vitest. ## Education B.A. Economics, University of Connecticut, 2020, with computer science coursework. ## Contact Contact form at https://jamesfeeney.dev/#contact. Email: hi@jamesfeeney.dev.