# James Dimachkie

> I’m James Dimachkie. My paid experience is in finance operations and audit, alongside independent AI-assisted research in evaluation and assurance. Based in Sacramento; open to Bay Area roles and to operations work at AI safety organisations.

This is my personal portfolio, not an employer website. I explain what I did, what the evidence supports and where its limits are. My billing improvements and enterprise-AI work are separate activities. Each case identifies whether it concerns employment, client work, an independent prototype or a fictional demonstration.

## Profile and contact

- [James Dimachkie](https://jamesdimachkie.com/): Professional background, selected work, applied-AI practice and contact information.
- [About James](https://jamesdimachkie.com/#about): Finance and operations experience, education, enterprise-AI support and independent projects.
- [Public CV](https://jamesdimachkie.com/assets/James_Dimachkie_CV_2026.pdf): One-page PDF connecting AI evaluation research with audit and finance operations, including contribution limits and career history.
- [Contact](https://jamesdimachkie.com/#contact): Email and contact form for roles and conversations about the work.

## Research and applied AI

- [Evaluator Assurance](https://jamesdimachkie.com/work/authority-chains): Public overview of Evaluator Assurance: Two Case Studies of Formal-Checker Integration. Private manuscript revised September 10, 2026; not submitted, published or peer reviewed. James set the research direction; Codex agents performed substantial design, implementation, analysis and writing. Human technical verification and external review remain outstanding. The manuscript and research archive are not public.

- [Receiver Reliance](https://jamesdimachkie.com/work/receiver-reliance): A 1,805-episode study of AI-agent handoffs, evidence and reliance decisions. Includes negative findings, limitations and links to the published papers.
- [Being checked is not the same as being right](https://jamesdimachkie.com/notes/being-checked): Why agreement between AI systems does not establish correctness, and what an independent check needs to challenge.
- [A system should earn its complexity](https://jamesdimachkie.com/notes/an-attention-budget): How useful instructions, context and verification can support AI-assisted work without accumulating redundant rules.

## Software and design case studies

- [Sandy Pines Nesting Place](https://jamesdimachkie.com/work/sandy-pines): Client website presenting five planned retreat stays, their differences and details still being developed.
- [Crewbrief](https://jamesdimachkie.com/work/crewbrief): A fictional interactive field-briefing demonstration connecting hazards, actions and crew handoffs. The case explains its scope and limits.
- [Catchwright](https://jamesdimachkie.com/work/catchwright): A fictional document-review demonstration separating proposed values, human confirmation and comparison with sample pay. The case explains its scope and limits.
- [Arbor & Eye](https://jamesdimachkie.com/work/arbor-eye): An independent, uncommissioned gallery prototype for twenty-eight watercolors, with individual artwork views and enquiry paths.
- [How this place came to be](https://jamesdimachkie.com/work/after-the-prompt): How I developed the illustrated portfolio with AI, including rejected directions and the optional scroll-driven film.

## Implementation notes

- [A clock-in needs its own ID](https://jamesdimachkie.com/notes/a-clock-in-needs-an-identity): An offline timeclock design decision: preserve entry identity through retries and recovery.
- [Eight fields before a comparison](https://jamesdimachkie.com/notes/eight-fields-before-a-comparison): How a small demonstration separates a proposed document value, a human decision and a rate comparison.

## Optional

- [Sitemap](https://jamesdimachkie.com/sitemap.xml): Canonical content URLs and their recorded modification dates.
- [Published research record](https://zenodo.org/records/22492561): The papers, frozen source, analysis outputs and reproduction materials discussed in the Receiver Reliance case study.

The HTML case studies are the primary explanatory pages. Standalone interactive demos use fictional data and are intentionally excluded from indexing. Research results describe the published study's bounded setting; they do not establish general AI safety or real-world security efficacy.