Skip to selected work
James Dimachkie
Menu

Finance operations & audit.

I’m James Dimachkie. My experience spans project coordination at BluSky Restoration, accounting support at The John Stewart Company and an audit internship at Deloitte.

I also direct independent, AI-assisted research into evaluation and agent oversight. The work below separates my contribution from the agents’ substantial assistance.

Read my CV (PDF)Explore Receiver Reliance

Sacramento, California · Open to Bay Area roles

jamesdimachkie@gmail.com

Further research: Evaluator Assurance. Two case studies of formal-checker integration. Private manuscript; not submitted or peer reviewed. Human technical verification and external review remain outstanding.

Read the public overview

Research & selected work

AI evaluation and agent oversight, followed by working software that illustrates my approach.

Receiver Reliance

Multi-agent oversight, published synthetic study

When should an AI agent rely on another’s work?
A decision at the receiving boundaryAn exact work product, its evidence, and the intended use feed a structural check. The host releases work after a pass; otherwise it may repair or hold. Answer correctness requires a separate check.Exact workproductEvidenceIntendeduseReceiving decisionRepair or holdRelease on passA decision at the receiving boundaryAn exact work product, its evidence, and the intended use feed a structural check. The host releases work after a pass; otherwise it may repair or hold. Answer correctness requires a separate check.Exact work productEvidenceIntended useReceiving decisionRepair or holdRelease on pass

The check can release a handoff, request one repair, or hold the receiver. The host supplies the facts and enforces the decision.

I studied when an AI agent should rely on another’s work. Across 1,805 synthetic episodes, a structural check cut improper handoffs by an estimated 1.0 per episode against transparent release, with an interval from −1.6 to −0.4. It also completed fewer seeded tasks. Wrong answers with coherent paperwork still passed.

My contribution & AI’s role
Me
I chose the research question, directed the study and decided to publish its costs and negative findings.
AI
Substantial assistance: specification, implementation, mathematical analysis, execution, review and writing.
Read the case studyRead the published study

Evaluator Assurance

Evaluator assurance, private manuscript

From an assurance claim to an accepted result
  1. 01 · ClaimWhat does the evaluation say it checked?
  2. 02 · EvidenceWhat supports that claim?
  3. 03 · DecisionDoes acceptance depend on that evidence?

A check’s existence and its influence on a result are separate questions.

Conceptual diagram of the research question. Private manuscript; human verification and external review remain outstanding.

I directed two AI-assisted case studies of formal-checker integration in evaluators. The local study and revised manuscript are complete; human verification and external review remain outstanding.

My contribution & AI’s role
Me
I proposed applying Receiver Reliance to the limits of public AI assurance claims, asked for rigorous testing and directed interpretation of the outputs.
AI
Codex agents performed substantial study development, implementation, execution, analysis, evidence coding, adjudication, review and manuscript drafting.
Read the research overview

Crewbrief

Workflow design, software in development

Crewbrief’s unresolved control state: two observations, zero controls selected, and a message that this category needs a control before continuing.

I directed this field-briefing demonstration to keep each hazard beside its control. No control means the briefing cannot continue. An affected-area hold stays in place after crew acknowledgment: reading the briefing does not clear the condition.

My contribution & AI’s role
Me
I directed LaborHounds and this demonstration.
AI
Implementation and review.
Read the case studyTry Crewbrief
Earlier software and design work

Catchwright — source-linked document review, retired product. Sandy Pines — a client website for a retreat in planning. Arbor & Eye — an uncommissioned gallery study.

James Dimachkie.

From audit and operations to AI assurance.

I want my work to help make advanced AI safer and more accountable.

I’m a Project Coordinator at BluSky Restoration in Sacramento, with responsibilities spanning project administration, branch finance and day-to-day operations.

AI systems assisted with the research, implementation, writing and review presented here, and generated the illustrations and film in the design archive. I chose the questions and directed the work; each case explains the division of contributions and the limits of its evidence.

AI tools in daily operations

Colleagues come to me for help with company-provided AI tools. I use them for drafting, proposals and supplier sourcing, and help others apply them to their work. I also support other offices with billing and collection difficulties.

Finance responsibilities and recognition

My work includes budgets, subcontracts, job costs, progress billing, work-in-progress reporting, customer payments, margin monitoring and weekly payroll. I maintain opportunity and contact records and support data migrations across business systems. I received a ROAR award for contribution.

Earlier work and education

Before BluSky, I supported Audit & Assurance work at Deloitte in Abu Dhabi and received recommendations for a full-time role. Family responsibilities brought me back to the United States; I later worked in property operations at The John Stewart Company. I have a BA in Administration with a finance concentration from California State University, San Bernardino. References and my Deloitte internship certificate are available on request.

Where I want to contribute

My primary direction is AI evaluation auditing and assurance, including research support, red-teaming and work that connects technical evidence to governance decisions.

I’m also open to finance and operations roles at organisations working on AI safety. That offers a direct way to contribute my existing experience while developing this research practice.

Read my CV (PDF)LinkedIn

Finding a source is only part of the search.

My local retrieval checks distinguish finding a relevant source from finding every source a question needs. Duplicate results cannot fill a missing part of the evidence.

Try a small evidence check

In this example, a returning assistant needs its saved identity and the current limits on system repairs. Either record is relevant, but the answer needs both.

What did the search return?

Returned: Identity record

Relevant source found
Yes
Required sources covered
1 of 2
Evidence for this query
Incomplete

The maintenance record is missing. A second copy of the identity record cannot supply it.

Illustrative inputs using the completeness rule in my local retrieval checks. This shows source coverage, not whether an AI’s answer is correct.

I explain more of my process in the three notes on method below.

Notes.

Decisions from the work,
with the reasons behind them and what changed.

A clock-in needs its own ID.

Review found that every queued offline clock-in in LaborHounds would fail the same way. I chose a device-generated entry ID over a new migration, so retries, reloads and rejected punches resolve against one entry.

On offline software
Read the note

Eight fields before a comparison.

I kept extracted values provisional in Catchwright until a person checked the source clause. Only confirmed hourly rates reach the pay comparison; a held hourly-rate field keeps the hourly comparison incomplete.

On source confirmation
Read the note

Also, on method: Six standards, and when each mattered, Being checked is not the same as being right and A system should earn its complexity.

Design archive: the earlier portfolio and optional film

Inside the atelier.

An earlier version of this site took the form of a paper workshop, with rooms for exploring, making, inspecting and remembering. I kept the original film here. It runs through nineteen seconds in four chapters as you scroll.

Optional, 8–15 MB. You can read the rest of the site without opening it.
How this place came to be: the drafts and the directions I did not keep
A cutaway paper workshop with rooms for exploring, making, inspecting, and remembering, connected by a red thread.
A frame from the film’s final reveal.

Let’s talk.

For research support, evaluation and assurance, or finance and operations at an AI safety organisation.

Read my CV (PDF)
jamesdimachkie@gmail.com
Send a message

Sent via Formspree. Your details are used to reply.

LinkedIn