Brian MillsAI Systems Engineer

AI systems engineer · agentic systems and AI for research and analysis

I build AI systems for complex analytical and operational work.

I design agentic systems that give AI bounded work and bring people in when a decision needs their judgment, alongside analytical methods that keep evidence, method, and uncertainty explicit. The right division of work depends on the task and its consequences. The two meet in one loop: agents execute inside a method, a check comes before the result, and a person sits at both ends.

Selected work ↓How it fits ↓

Experience

  1. Jun 2026 – present

    Inside Success / ISTV

    Organizational AI & agentic systems

    Building shared company knowledge, person-specific and organizational agents, proactive workflows, and bounded autonomous work across real tools. Includes an evaluation of agent answers: every citation must reopen to the exact source passage, and the agent must decline to answer when evidence is absent, tested with negative controls and a human-reviewed reference set.

  2. Nov 2025 – Jun 2026

    Steno

    AI digital twins & model fidelity

    Persona and behavior modeling, fine-tuning pipelines, and evaluation of whether generated behavior preserved characteristics of the source person. Found that the AI grader varied more between runs than the effects being studied, stopped drawing conclusions, switched to a deterministic grader, and rescored earlier experiments.

  3. Sep 2020 – Oct 2025

    RAND Corporation

    Policy Analyst & Doctoral Fellow

    Built knowledge graphs from documents with LLM extraction (defense co-production with NATO allies; the full U.S. Code) and presented the method to RAND researchers in 2024. Researched how institutions adopt AI and how AI can manipulate people, including the methodology for RAND’s study of its own AI use. Turned social media at massive scale into evidence, including the full Reddit archive (about 6 billion comments) and, as co-author of RAND’s Public Perceptions of 5G Technologies (2024) for the Department of Homeland Security, its conspiracy-theory analysis.

How it fits together

Engineering autonomy meets analytical rigor.

Engineering sets how work moves: planning, delegation, context, verification, escalation. The analytical layer sets what counts as good reasoning: evidence, method, uncertainty, justified conclusions.

The human–AI loop: two pillars, one loop A person sets direction. Context flows down into the agentic engineering layer, where agents execute the work. Inside that layer, an analytical method constrains the reasoning. The work leaves the engineering layer through an independent check and becomes a checked result, which returns up to the person, who decides. The decision feeds the next round of direction. Rarely, the engineering layer escalates a judgment call straight up to the person. What was proven loops back into the agents for reuse. next round Person sets direction Person decides 01Agentic engineeringmoves the work Agents execute 02Analytical method constrains the reasoning check checked result
Context

Context & representation

Give each actor the smallest useful view.

Agents receive bounded context for their role; humans get task-appropriate review artifacts rather than raw transcripts, giant diffs, or one universal dashboard.

Proves: I treat context and representation as architectural controls.
Escalation

Attention-aware orchestration

Escalate judgment, not activity.

An orchestrator manages work, context, budgets, retries, and recovery. Routine execution stays quiet; only decisions that genuinely need human judgment become attention requests.

Proves: I design control planes for autonomous work, not just individual agents.
Verification

Verification & feedback

Make failure visible, reversible, and useful.

Independent verification, explicit evidence states, recoverable execution, and feedback into policies let the system improve without silently drifting.

Proves: I build learning loops around agent behavior rather than trusting one-shot generation. See a recorded verification run ↗
Reuse

Composable capabilities

Make each project improve the next one.

Reusable primitives and evidence-backed capabilities let agents inspect what already works, compose proven behavior, and implement only the gaps instead of rebuilding from scratch.

Proves: I optimize for compounding system capability, not local project speed alone.
  • person
  • AI / engineering layer
  • analytical method
  • independent check
  • result
  • rare path

Current direction: build more end-to-end demonstrations where these layers work together, rather than forcing every existing project into one artificial runtime.

Proof

Systems that show it.

Separate systems, not one invented mega-architecture. Together they show the engineering and analytical sides of the same human–AI design.

Writing

How I think about the work.

Essay

Automating Knowledge Work Is an Architecture Problem, Not a Prompting Problem

Why reliable knowledge-work systems need explicit evidence, state, methods, models, and evaluation—not just increasingly elaborate prompts.
Read ↗

Interactive essay

Keeping company AI agents from leaking secrets or spending without approval

Two checks in the program around the AI, tried on an invented company: a written rule leaked hidden facts in 5 of 50 test answers, and a check that kept those records from the AI leaked none.
Read ↗

Interactive essay

Translating an app's data design into a database can silently drop its rules

One design, three databases, and data built to break each rule: TypeDB stored a duplicate that PostgreSQL and MongoDB refused.
Read ↗

Contact

Interested in research engineering, applied AI, agentic systems, and AI for complex analytical work.

Email meLinkedInGitHub