Remote, U.S. Eastern timeRemote, U.S. Eastern
Content engineer & GenAI enablement lead with 12 years in content
Editorial Lead at Credit Karma. I design human-in-the-loop AI systems for editors and run the GEO program that gets our pages cited by AI answer engines.
Open to full-time senior IC roles in content engineering, GenAI enablement and GEO. Fully remote · U.S. citizen, Eastern time
A full-time, senior individual-contributor role where content engineering, GenAI enablement or GEO is the core of the job.
I design AI systems that check content against editorial standards before an editor reviews it.
Continued in SystemsI write the prompts, evaluations and rubrics that tell a team when AI output is ready to ship.
Continued in WorkI shape pages so AI answer engines cite them and search engines rank them.
Continued in GEOI put short lessons inside the product, so people understand a decision before they’re asked to make it.
Continued in WorkGEO · Credit Karma
Credit Karma trailed that competitor in AI presence and in citations across the answer engines people now ask first. I ran the GEO program and designed the analysis tooling behind it, which fuses search rankings, AI-citation data and content metadata and showed us which pages the models were lifting and why. We shipped the page changes it pointed to, and in under a year we were at parity on both measures.
Content engineering · Credit Karma
Keeping our article library current meant a manual review of every article, one at a time. I rebuilt the process as an AI evaluation pipeline: it checks every claim against a primary source, quotes the source verbatim, tests every link live, scores the piece against our internal standards and SLAs, and writes the ticket. Reviews that took hours now take minutes, and an editor still signs off on every one.
LLM evaluation · Credit Karma
Intuit Assist answered members’ questions from our educational library, and I led the evaluation framework that decided when it was ready for them. Every answer was scored against a written rubric for accuracy, hallucination, honesty, trustworthiness and brand voice, and the assistant had to clear the bar before launch. Our editors were the subject-matter experts: they validated answers against current, legally approved source articles in our library.
Selected agents from the multi-skill GenAI platform I architected at Credit Karma. For each one I wrote the rubric first, then directed AI to build the tooling against it.
A rubric-driven audit agent that reviews an article the way a senior editor would, then writes the ticket. It scores each piece against our internal standards and SLAs, checks claims against primary sources, and outputs a severity-ranked ticket: what’s wrong, why, the primary-link sourcing, legal substantiation, and SEO and GEO recommendations. Once the CMS change goes live, it updates the team’s tracking database on its own.
Result: Every article in the library is reviewed against the same written standards, and an editor accepts or rejects every ticket.
The source-checking layer that sits on top of the audit agent. It verifies claims only against a predetermined allowlist of approved sources, returns verbatim citations, and runs live link checks, so “fact-checked” means checked against something a lawyer would accept.
Result: Fact-checking happens claim by claim across the whole library, and every claim traces back to an approved source.
Crawls a page and recommends generative-engine-optimization changes based on how LLMs actually fetch and lift content, current GEO practice, and what competitors are doing, then drafts first-pass revisions for an editor to accept or reject.
Result: Editors start from a drafted revision instead of a list of general GEO advice.
Fuses keyword-research data with AI-citation tracking to recommend long-tail query and keyword optimizations that hold classic search relevance while improving AI-citation share against competitors.
Result: Search rank and AI citation sit in one view, instead of two dashboards that disagree.
Takes an article URL plus optimization instructions and produces a first-draft, freelancer-facing brief ready for review and send: the ask, the sources, the structure, the standards.
Result: Briefs editors used to write from scratch now start as a draft ready for review.
Reads card-news updates as they arrive by email and drafts the legally compliant CMS updates for review and publish.
Result: Compliance updates become review work instead of writing work, and an editor approves every draft before it publishes.
The shared design principle
Every one of these systems drafts, and a person decides. Anything that acts outside the system waits behind an approval gate, and the standards each one enforces are written down as rubrics an editor can read and a model can check. When I say human-in-the-loop, I mean the checks run before an editor reviews the work, and the editor makes the final call.
Five pieces of work, told the way I’d tell them in a first call: the problem, what I did, and what happened.
AI strategy · Credit Karma
For the fiscal year I wrote the editorial org’s operating model for AI plus HI, artificial intelligence plus human intelligence: a detailed plan to automate a large share of editorial’s manual work so that time moves to strategic work.
LLM evaluation · Credit Karma
Intuit Assist was Credit Karma’s first proprietary chatbot. Before it launched, someone had to decide when its answers were good enough to put in front of members.
Education as a product · Credit Karma
Many members came to Credit Karma with little or no credit history and not much financial vocabulary. So we taught inside the product instead of linking out to articles: on the app’s front door, and in a short financial-literacy curriculum for Gen Z members across the insurance, credit card and personal loan tabs.
Experimentation · Credit Karma
Marketing’s push-to-shop program sent a notification that dropped members straight into the credit card shop. With the credit cards team I added a step in between: the tap landed on education first, and a call to action then took people into the shop.
Data-driven SEO · Credit Karma
Car insurance is priced state by state and moves with your credit, which is hard to explain in one article. So in 2024 I led a series with one article per state, written by our freelance writers and cross-referenced by credit tier, so a reader could see what someone with their credit pays where they live.
Generative engine optimization is the work of getting a page cited by AI answer engines as well as ranked by search. These are the practices I’ve arrived at by running a GEO program at scale and by testing them on my own writing.
Answer in the first 60 words. The question the headline implies gets answered in the opening paragraph with a date, a number or a name. Answer engines lift that paragraph, so it has to be able to stand alone.
Give the models something to cite. The foundational GEO paper (Aggarwal et al., KDD 2024 (opens in a new tab)) found that adding statistics, quotations and cited sources raised a page’s visibility in generative-engine answers by up to 40% on its benchmark, while keyword stuffing hurt it. Three or more attributed hard facts per piece is the floor.
Bring one thing nobody else has. Proprietary data, first-hand observation, your own numbers. It’s the part a generic site can’t copy and the part a model has no other source for.
Quote a primary source once, verbatim. One direct quotation from the official notice, the filing or the spec, in quotation marks, with the source linked.
Spell entities the official way and repeat the main one. Product names, codes and proper nouns exactly as the authority writes them, in the H1, the first paragraph, one H2 and the meta description.
Write each heading as the question a person would type. Phrase it the way someone would ask an LLM, and answer it in the first sentence underneath. Each section should work as a 40–90-word standalone answer before the detail starts.
Put the trust signals in the plumbing. A short FAQ rendered as FAQPage schema, a sources list rendered as citations, and a named author rendered as Person schema, so a crawler can read the experience and expertise the prose describes.
Keep it fresh in place. Update the existing piece when the story moves, with a visible modified date, instead of publishing a near-duplicate that splits the signal.
One narrow topic per piece. A 400–850-word page that answers one question beats a 2,000-word guide, for readers and for citation. Write the siblings as separate pieces and link them with descriptive anchor text.
Keep the classic SEO basics. Canonical URLs, a sitemap with last-modified dates, fast pages, titles under 65 characters, descriptions under 155, and a robots policy that lets the AI crawlers in. An llms.txt file is a cheap, voluntary extra.
Write like a person. Attribute once per cluster of facts, then talk. No throat-clearing openers, no stacked hedges, no list of ten generic tips, and one real opinion per piece, labelled as a read.
Measure it the way you’d measure rank. A fixed panel of buyer-style prompts run monthly across ChatGPT, Perplexity, Gemini and Google’s AI modes, logging which domains get cited, against a numeric target. Before writing, run the question yourself and beat the top cited pages on specifics.
Published work with my byline or my edit on it, each labelled for what I did: written, edited, or updated (an existing piece I refreshed). Search positions are from a web-search check on 13 October 2026.
Clipped from the live pages on 14 October 2026: the headline, the credit line, the date and, where the page has one, the lead photo, printed in one ink.
74 pieces from 2015 to 2026: 45 at Credit Karma, 28 at BestReviews and 1 at The Wenatchee World. Earlier newspaper bylines: 156 at The Wenatchee World from 2014 to 2018 and 88 at The Spectator (UW–Eau Claire) from 2012 to 2013.
I started out running editorial workflows by hand, and now I automate them.
I own GenAI enablement for the editorial org: the multi-skill GenAI platform the team uses daily, the GEO program, and AI evaluation for content QA.
I was a senior editor and the editorial org’s AI strategy lead.
I owned the SEO content pipeline for my verticals end to end, with its maintenance, compliance and publishing workflows.
Liveliquidpixelz.io
It’s a live-commerce shop for One Piece trading cards, with a news page that publishes a new article every day.
I built it and run it with a crew of AI agents I direct. Claude Code writes the code to my spec, and I test what ships.
Visit liquidpixelz.io (opens in a new tab)The GEO practices its news agents follow
The vault (opens in a new tab), where each graded card is listed with its grading details.
davidheiling.com
I built it the same way: Claude Code writes the code to my spec, and it runs on Next.js and Vercel.
For twelve years I’ve worked in content. I started as a sports reporter at a daily paper in Washington state and became its sports editor by leading the paper’s move to digital. Since then I’ve run a hyperlocal newsroom’s workflow, built the production system for an e-commerce content operation, and spent four-plus years at Credit Karma, where I’m now the Editorial Lead.
I architect the systems, direct the AI that builds the tooling, and use those agents every working day. Claude Code writes the code to my spec and rubric, and I test what it builds against real work. What I bring is editorial judgment, written down as standards a model can check before an editor reviews anything. The blueprint was the pipeline I used to run by hand as an editor: keyword research, a monthly budget, assignments, two edits, publish.
I was the first in my family to graduate college, where I ran the student paper. At 25 I booked a one-way ticket to Croatia and spent five years working remotely across 60-plus countries; India and Kenya are still my favorite places. Away from work I’m with family, hiking in the forest, meditating or collecting, and I’m a dog and cat person.
I also run a small live-commerce business with a crew of AI agents I built. It’s my personal AI lab, where I test agent patterns before I bring them to a team.
Highlighted
The MBA gave me the business side of content: what it costs, what it returns, and how to run the project that ships it. I use it whenever I work between editors, product managers and finance.
First in my family to graduate college.
The methods and tools I work with, grouped by lane.
My lanes are content engineering, GenAI enablement and GEO, and I’m also open to SEO, content strategy and product marketing content. Industry is open: any company that needs AI working for its content is a good fit.
If you’re hiring in one of those lanes, or you want to compare notes on the work, email me. I read everything and reply in writing. A sentence or two about what you have in mind is plenty.