Field notes / A continuing education

Learning that changes
how I build.

I’m moving deeper into research engineering for reliable ML and applied AI. These are the ideas I’m carrying from production data systems into that work, and the questions I want to investigate next.

This is a learning record, not a publication list. The notes below synthesize my existing technical study. Research questions describe the direction of my work, not completed findings.

From the current investigations

Field notes

TRACE / Graph systems

Before improving the ranker, inspect the graph.

What dependency analysis is teaching me about separating missing evidence from poor decisions.

Diabeteris / Data & evaluation

The research system starts before model training.

Why provenance, data availability, and experimental boundaries belong in the first version.

AXIOMERIS / Research practice

Shared methods. Specific assumptions.

A working approach to connecting research domains without pretending their problems are identical.

Method / How I want to work

A question, a baseline,
and a record of what changed.

The research habit I want to strengthen is simple: make an idea specific enough to test, then leave enough context for someone else to challenge the result.

  1. 01

    Frame

    Write the question, the assumptions, and what would count as a useful result.

  2. 02

    Reproduce

    Establish a baseline and record the data, configuration, and evaluation protocol.

  3. 03

    Investigate

    Change one thing deliberately. Study the errors as closely as the headline result.

  4. 04

    Communicate

    Write down the evidence, limitations, and the next question the experiment creates.

Notes / Technical foundations

Five lessons I keep returning to.

View the credentials

01Data engineering

Start with the question the data must answer.

Partitioning, file formats, and orchestration shape what a team can reliably ask of its data. The lesson I carry from AWS data engineering and Foundry is to design around access patterns and provenance together. A fast query is useful; a fast query whose inputs can be explained is a better foundation for research.

In practice. Document the source, the transformation, and the expected consumer. Make reruns safe before making them fast.

AWS Data Engineer · Palantir Foundry

02Model evaluation

A metric has to answer a real decision.

Feature engineering and model selection only make sense alongside a clear evaluation target. My machine learning study pushed me to question what an aggregate score hides: class imbalance, different error costs, and performance that changes across populations. The useful question is what happens when the system is wrong.

In practice. State the decision, choose a baseline, examine failure cases, and make the limits of the evaluation visible.

AWS Machine Learning · Specialty

03ML systems

Reproducibility belongs in the architecture.

A training run is only one event in a model’s life. Data versions, feature definitions, deployment behavior, and monitoring determine whether that result can be trusted later. MLOps made the relationship clear: reproducibility and rollback are design requirements, with consequences for every stage of a pipeline.

In practice. Keep configuration and data lineage with the result. Define evaluation gates and a recovery path before deployment.

AWS ML Engineer · MLOps Microcredential

04LLMs & agents

Evidence should travel with the answer.

Tool use and longer context do not remove the need for boundaries. My study of agentic systems and Claude architecture sharpened how I think about permissions, context, and grounding. A capable agent still needs a clear account of which evidence it used, which tools it can call, and when a person must decide.

In practice. Keep context auditable. Evaluate grounding separately from fluency, and give tools the narrowest useful scope.

Agentic AI · Claude Architect · Accenture

05Technical judgment

Write the constraint before the conclusion.

Reviewing assessment items requires precision: a choice has to be defensible against the stated conditions. That discipline transfers to architecture and research writing. When a recommendation changes with scale, cost, or risk, those assumptions belong next to the recommendation.

In practice. Explain the alternatives, cite the basis for a claim, and distinguish what was observed from what was inferred.

AWS Subject Matter Expert · Professional & Associate

Research direction / Open questions

The next layer of understanding.

Evaluation

When does a better benchmark score translate into a better decision?

Baselines, error analysis, distribution shift, and the cost of false confidence.

Data & provenance

How much of a model’s behavior can we trace back to its inputs?

Data lineage, feature definitions, graph structure, and reproducible experiments.

Applied AI

What evidence should an AI system provide before someone acts on its output?

Retrieval quality, grounded generation, explicit uncertainty, and human review.

Read the journal Explore the systems