Dylan Davis

Dylan Davis builds AI systems and automation and coaches business owners and executives to build tools around their own workflows. He is affiliated with the Gradient Labs consultancy and publishes technical explanations on D-Squared, including a three-GPT executive-summary workflow he created and a survey of AI-assisted vulnerability patching.[2][3][5][4]

Davis’s executive AI coaching centers on a practical task and the client’s real files. His format is direct: “No curriculum. No slides. You share your screen, I guide, you drive.” In two-hour working sessions, he explains the build as it happens; clients leave with a working tool, a recording and transcript, and follow-up lessons and homework. His stated aim is for clients to learn enough to build the next tool themselves.[3]

The coaching page describes Ben McPeak, president of Vertical Jet Sales, moving from manually gathering aircraft data and preparing market reports to completing a market analysis after one session. McPeak’s testimonial reports: “What would have taken me an entire day took 45 minutes.” The page also credits work built together with John Tesh, including turning decades of radio archives into a content engine and enabling the broadcaster to run Codex and Claude himself. These are client testimonials and first-party accounts of the collaborations.[3]

Before this coaching work, Davis says he built AI systems and automation inside Mastercard, HSBC, and Bank of America. His earlier public writing provides concrete examples of his approach to dividing AI tasks and retaining human review.[3][5][4]

In July 2024, Davis published Three GPTs Walk into a Bar and Write an Exec Summary, describing three Custom GPTs he created for frequent or time-sensitive summary requests. Insight Extractor selects ten distinctive report insights; Exec Sum Writer takes those insights and the full report to draft four paragraphs; Revisor Bot edits the draft for brevity, clarity, and readability. The user passes each stage’s output to the next.[5][4]

Davis chose three specialized GPTs because he favored breaking a complex task into smaller subtasks rather than assigning everything to one prompt. He recommended checking the final summary for hallucinations and editing it for the writer’s tone. He also qualified the workflow’s purpose: it assists with time-constrained requests, while writing remains a way to think. As he put it, “Remember, improved writing leads to improved thought, so don’t always default to using LLMs for your writing endeavors.”[5]

Davis’s Self Healing Code, published November 3, 2024 and subsequently updated, explains other teams’ work on automated vulnerability patching. His recommendations emphasize focused code context, conventional security tools, and prompts assembled for the particular vulnerability. “Don't throw out traditional security tools,” he advises, treating LLMs as additions to static analysis, testing, and CI/CD pipelines.[4]

His LLMPatch explanation makes that approach concrete: Joern produces a program dependence graph to identify related code; the system retrieves examples with similar root causes, generates five candidate fixes, and uses LLM reviewers before final human validation. Davis distinguishes two review questions—whether a patch preserves the original functionality and whether it fixes the security problem. His contribution is explaining and assessing the researchers’ system.[6][4]

Davis also appeared on Latent Space in Anthropic vs Cognition on Multi-Agents: A Breakdown with Dylan Davis.[1]