Applied AI Work Samples

Daniel Taylor

I use AI as part of a disciplined problem-solving workflow: frame ambiguous problems, compare and evaluate outputs, structure research, test hypotheses, identify failure modes, and convert the result into something usable.

AI EvaluationAI QA / Model QualityResearch & SynthesisSystems AnalysisAI Workflow Design

Human judgment at the center

I am most useful where the problem is not fully specified and the output has to be judged, corrected, or organized rather than merely generated. I use AI for breadth, iteration, comparison, drafting, and tool assistance; I retain responsibility for framing, validation, error detection, and final selection.

This portfolio is deliberately transparent about that division of labor. The work samples below are examples of human–AI collaborative problem solving, not claims that AI was absent from the process.

Working principle

Use AI to expand the possibility space and reduce mechanical workload. Use human judgment to determine what is true, relevant, coherent, and worth implementing.

1. FRAME
2. EXPAND
3. TEST
4. JUDGE
5. IMPLEMENT

Selected work samples

Case 01

LLM Response Evaluation & Reasoning Quality

Evaluation practice set
Problem

Compare competing AI answers, identify the stronger response, and explain the decision clearly without over-writing the rationale.

What I focused on
  • factual correctness versus plausible-sounding claims
  • whether evidence actually supports the conclusion
  • unwarranted certainty and premature diagnosis
  • single-point-of-failure reasoning and practical risk
  • concise explanations that name the decisive issue
comparative evaluationlogical validityuncertainty calibrationclear rationale
Read full case study →
Evaluator-style examples
“Response B is better because a fan making noise doesn't prove that it is broken. Response A jumps from an observation to a diagnosis without ruling out other plausible causes.”
“For most people, one external SSD may be sufficient in ordinary use, but it is still a single point of failure. Multiple backups are the stronger recommendation.”

The goal is not maximal explanation. It is to identify the controlling error and state it in a form another reviewer can audit.

Case 02

Iterative Technical Diagnosis: Fedora, VPN, Browser Networking

Resolved troubleshooting case
Problem

A VPN connection on a Fedora laptop was unreliable or slow. Several layers could plausibly be responsible: IPv6 routing, system proxy behavior, VPN configuration, or browser DNS handling.

Diagnostic loop
Observe failure→Test IPv6 hypothesis→Reject→Test proxy→Connection works but slow→Isolate browser DNS

The key step was treating each result as evidence that updates the model rather than continuing to defend the first hypothesis.

Outcome

Disabling Brave's Secure DNS feature removed the slowdown and restored normal page loading. The earlier IPv6 change had not fixed the issue; enabling the system proxy changed the failure into a working-but-slow connection and became an important intermediate clue.

What this demonstrates
  • layered systems reasoning
  • hypothesis testing
  • using negative results productively
  • isolating interacting components
  • updating the working model as new evidence arrives
Linuxnetworkingdiagnostic reasoninghuman-AI debugging
Read full case study →
Case 03

Research Architecture: Making a Complex Technical Field Legible

Research / documentation sample
Challenge

Build a rigorous, non-specialist explanation of cybernetics without reducing it to vague “systems thinking,” burying the reader in mathematics, or allowing adjacent concepts to blur together.

Architecture
  • define the reader outcome before drafting
  • identify central questions and required concepts
  • maintain explicit distinctions: feedback vs. command, regulation vs. domination, homeostasis vs. adaptation, system vs. environment
  • specify exclusions to prevent scope drift
  • use examples as tests of conceptual clarity
  • set chapter-level success criteria before prose generation
Why it matters

This is the same underlying task found in technical documentation, policy analysis, knowledge-base design, and AI-assisted research: preserve the structure of a difficult domain while translating it for a different audience.

Quality controls
Preserve conceptual integrity. Prevent false equivalence. Distinguish informative error from simple failure. Make the structure readable without flattening the subject.
research synthesisinformation architecturescope controltechnical communication
Read full case study →
Case 04

AI Production Workflow Design for Long-Form Research

Workflow / process design
Problem

Large AI-assisted writing projects drift when generation begins before structure, source requirements, exclusions, and quality controls are stable. The solution was to design a gated production system rather than ask for “a book” in one pass.

Six-phase workflow
Canonical structure→Chapter skeletons→Core-ideas expansion→Drafting→Revision / alignment→Supporting artifacts
Control mechanisms
  • no prose drafting before the skeleton is approved
  • each section has purpose, claims, examples, exclusions, and tone constraints
  • revision explicitly checks drift, redundancy, overstatement, inconsistency, and conceptual bleed
  • source files are treated as canonical inputs rather than regenerated from memory

This sample demonstrates workflow orchestration: designing the process through which AI contributes, not merely writing prompts for individual outputs.

workflow designAI orchestrationquality gatesspecification writing
Read full case study →

Capabilities demonstrated across the work

Evaluation

Compare model outputs, detect unsupported conclusions, identify reasoning defects, apply criteria consistently, and write concise rationales.

Research

Decompose broad questions, separate evidence from interpretation, maintain competing models, organize sources, and synthesize findings into usable structures.

Systems analysis

Model interactions, trace failure across layers, identify boundaries and dependencies, and revise hypotheses when observations conflict with the current explanation.

Workflow design

Turn open-ended work into staged processes with inputs, decision points, exclusions, quality gates, and explicit completion criteria.

AI collaboration

Use AI for rapid iteration and breadth while maintaining human ownership of framing, verification, correction, and final judgment.

Technical communication

Translate complex material into clear explanations, specifications, decision logic, and reusable reference structures without losing important distinctions.

Best-fit work

Immediate fit

  • AI response evaluator / human rater
  • LLM QA and model-quality review
  • rubric-based annotation and audit
  • research verification / fact checking
  • AI-assisted technical or research documentation

Growth path

  • AI QA test design and failure analysis
  • AI research / systems analyst
  • AI business-systems or workflow analyst
  • AI enablement / workflow orchestration
  • small, high-leverage AI-assisted project work
Positioning

I am not presenting myself as a machine-learning engineer. My strongest value is at the interface between a difficult problem and an AI system: clarifying the problem, structuring the work, judging output quality, catching failures, and turning the result into a coherent deliverable.