J·C·Fjelstul
Consulting LLC

The hardest part of an ML project isn’t the model.

It’s asking the right questions.
I bridge the gap between rigorous science and applied machine learning. I help organizations move from messy, unstructured text to production-grade systems by asking the right questions, carefully framing the business problem, and building domain-tailored, cutting-edge solutions.
Schedule a consultation
Work featured by
The Washington Post
The BBC
The Times
FiveThirtyEight
Sky Sports
The Markup
Services

ML Strategy

From readiness assessments to tailored strategy and training workshops, I can help your organization develop and implement an AI/ML strategy that advances your business objectives. Invest in AI/ML capabilities knowing you have a clear strategy.

Typical Investment

From $5,000

Details

ML Systems

I build custom production-ready AI/ML systems trained on your data and tailored to your substantive domain. I build classification systems, information extraction systems, and recommender systems that work because they have been specifically trained to understand the specialized language of your specific field.

Typical Investment

From $20,000

Details

Data Visualization

I design publication-quality visualizations and interactive data features that tell a compelling story using data. I implement best-practices from data journalism — every design decision serves the argument, not just an aesthetic.

Typical Investment

From $1,000

Details

The Differentiator

Science + Engineering
Most ML projects don't fail because of the model. They fail because the problem was framed wrong, the labels were poorly defined, or the evaluation metric didn't reflect what actually mattered in production.
Fixing those problems requires a different kind of expertise — not more advanced architecture or more compute, but better methodology. Asking the right questions. Carefully framing the problem. Thoughtfully operationalizing key concepts. Selecting appropriate methods, not trendy ones. That work happens before a single model is trained.
I have extensive experience conducting quantitative research at leading universities, working with some of the most complex text there is: dense legal documents, multilingual court judgments, and technical regulatory documents, where a misclassification isn't just a metric — it's a wrong answer about what the law says. That environment trains a specific skill set: defining valid labels, anticipating measurement error, and determining whether a model has actually learned the concept you care about.
I bring research-level rigor to every AI/ML project — from our first conversation about how to frame the problem to the final handoff to your team.
Recommended posts

Not Every NLP Problem Needs a Frontier Model

Frontier LLMs are capable. They're also expensive, slow, and prone to hallucination. For many NLP tasks, using a fine-tuned BERT model will be more accurate, easier to audit, better suited to your domain, and cheaper by orders of magnitude — and it'll keep your data off someone else's servers.
Read

Making a Monolingual Model Bilingual with Domain Adaptation

You have an English BERT model that works well on legal text. But your corpus is bilingual. Here's how domain adaptation on a bilingual corpus can produce a model with strong masked language modeling performance in both languages — and why legal text makes this work better than you might expect.
Read

Domain Adaptation or Fine Tuning?

Fine-tuning and domain adaptation are often used interchangeably, but they solve different problems and require different approaches. Getting the distinction wrong is one of the more expensive mistakes in applied NLP.
Read
Ready to build?

From Problem Framing to Production.

Whether you need a domain-adapted text classification model, or an end-to-end recommender system with RAG, I help you ask the right questions, frame your business problem, and build cutting-edge AI/ML solutions.

Schedule a consultation