Work with me · AI engineering & technical advisory

Clarity for your next AI decision.

I help engineering teams evaluate RAG applications and AI agents, understand their tradeoffs, and make decisions grounded in evidence. Bring a workflow, an observed problem, or a decision your team needs to make.

Discuss your project Explore the options

A few lines about your system and the question you are working through are enough to start.

Choose a starting point

Three ways to work together

A focused starting point

AI reliability diagnostic

Your application gives plausible answers, but you need to understand which ones are supported and where the failures begin.

What you receive

Error taxonomy and reusable evaluation rubric.
Review of retrieval, evidence, and memory.
Written report and a 60-minute team review.

One workflow · typically 20–30 cases · about two weeks. Timing and price agreed before starting.

Discuss a diagnostic

A decision before more complexity

Architecture & resource review

You are adding retrieval, memory, or agent steps and need to understand whether the extra complexity earns its cost.

What you receive

Retrieval, memory, inference, and MLOps review.
One agreed baseline or alternative comparison.
Tradeoffs and prioritized engineering recommendations.

One workflow and one comparison. Quality, latency, and cost examined where records support them.

Discuss an architecture review

Support as your system evolves

Ongoing technical guidance

Your engineers need a regular sounding board for evaluation, experiments, and the decisions that shape an evolving AI system.

What you receive

Evaluation standards and reusable review criteria.
Experiment design and interpretation of results.
Architecture reviews as your system evolves.

An agreed cadence, scope, and price. Implementation can be discussed separately.

Discuss technical guidance

From first conversation to a useful decision

A straightforward process

Start with the problem
Send a short description of your workflow, one example or concern, and the decision you want to make. We use that to assess the fit.
Agree on the work
Before starting, we define the scope, deliverables, price, timing, and appropriate access to the material we need.
Leave with a decision
We review the findings together, distinguish evidence from uncertainty, and identify which change or experiment deserves your team's attention.

Examine the approach

Public work you can inspect

Memory & evaluation

Memory Is Not Context

A research case examining resource accounting, memory strategies, and a verifier that can still be wrong.

MLOps & release workflows

Prompts Are Release Artifacts

A proposed release workflow for prompts, explored through a synthetic documentation example.

Music, embeddings, and narrative arcs

My independent music research explores how ideas recur and change across songs, albums, and a band’s discography. I use embeddings and NLP to study relationships between lyrics, then return to listening and close reading to examine what those relationships mean. I also study narrative arcs in literature: how a sequence develops an idea, and how a later passage changes our interpretation of an earlier one.

These questions shape my experiments on representation, memory, and verification. What does a representation preserve? Which detail must a system retrieve to support an interpretation? Where does a confident reading outrun the evidence?

The public work linked here consists of research studies and educational examples, with their own methods and limitations. They are not client results. Applying a finding to your system requires its own evaluation.

Before we start

How is pricing agreed?

I quote each engagement after we define the question, scope, deliverables, and access requirements. We agree on the price before the work starts; a diagnostic, architecture review, and ongoing guidance have their own scopes.

What should I prepare or share?

Start with a non-sensitive description and an example you are authorized to share. If we proceed, we agree on the representative cases, traces, outputs, and source material required, along with appropriate access.

Does this include implementation?

These engagements focus on evaluation, review, and technical decisions. Production integrations, large-scale data preparation, and implementation work need a separate agreed scope. Improvements in quality or cost are outcomes to measure.

Three ways to work together

AI reliability diagnostic

Architecture & resource review

Ongoing technical guidance

A straightforward process

Start with the problem

Agree on the work

Leave with a decision

Public work you can inspect

Memory Is Not Context ↗

Prompts Are Release Artifacts ↗

Music, embeddings, and narrative arcs

Before we start

Memory Is Not Context

Prompts Are Release Artifacts