Consulting · Kelly Enterprises LLC

Work with me

I take on applied machine learning and agentic systems work — usually where the problem sits between disciplines, and usually where the hard part turns out not to be the model. Small number of engagements at a time, so the fit matters.

What I do

  • 01

    Owning a platform end to end

    Application, data infrastructure, cloud and hybrid systems, integrations, operations, and security — as one system rather than six handoffs. I still spend most of my week in the code, which is the only way I know to keep architecture honest.

  • 02

    Document intelligence and extraction

    Agent-driven pipelines applying OCR, NER, and structured extraction to messy real-world corpora — scanned records, semi-structured PDFs — and mapping them into schemas you can actually query and audit.

  • 03

    Regulated, auditable data systems

    Translating regulatory guidance and scientific literature into domain models that hold up under audit. When the rules change mid-quarter, the pipeline has to change with them without losing traceability of what was reported and why.

  • 04

    Applied ML that survives real data

    Forecasting, computer vision, and evaluation on proprietary data that is partial, mislabeled, and disagrees with its own documentation. Most of the work is the pipeline and the error analysis, not the architecture.

How engagements run

How I work

  • Structure the knowledge before reasoning over it

    Most AI projects stall because the corpus is sediment — six versions of a policy, four superseded, none saying so. No model quality resolves a contradiction nothing was assigned to notice.

  • Traceable beats impressive

    An answer you cannot follow back is an answer you cannot debug, and a system you cannot debug is quietly abandoned two quarters after launch. I build for the audit, not for the demo.

  • Modularity over rebuilding

    Techniques transfer. What one industry treats as an open research problem is frequently a solved commodity in another. A lot of my value is knowing which building to look in.

  • I'll tell you when it isn't a modeling problem

    Most clients arrive convinced they need a better model. Usually they need better data plumbing, and saying so costs me the bigger engagement. I'd rather be right.

Start with the problem, not the solution

Tell me what's actually going wrong and what you've already tried. That's a far more useful first message than a spec — and it's the fastest way for me to tell you whether I'm the right person.