More AI work than hands?
Senior help, by the hour.

What I take on

The pieces a busy AI team falls behind on.

Each is work I have designed and delivered, for enterprise clients or in public code you can read.

• AI agents and tool use

  An agent that reads, decides, and calls the tools your team already has: retrieval, SQL, web search, your own APIs. Single-agent designs for production work, and multi-agent where the problem genuinely calls for it.

• Retrieval and plain-English data access

  RAG over your own documents with the source shown on every answer, and NL2SQL so people can ask the warehouse a question without waiting on an analyst. Two different tools, built and tested separately.

• Model evaluation and right-sizing

  Which model is good enough for this job, and what the cheaper one costs you in quality. Your own workload run through the candidates, graded blind, with the cost line next to the score.

• Machine learning on your own history

  Forecasting, classification, and early-warning models in Python and scikit-learn, validated on data the model never saw before anyone relies on it.

• Open-weight models on GPU

  Fine-tuning and deploying an open-source LLM on your own GPUs, for when per-token pricing or data handling makes a hosted model the wrong fit.

• Scoping the pilot

  Discovery, architecture, and an effort and ROI estimate for the AI work on your list, so the backlog gets worked in the order that pays.

Work you can read

Public code and write-ups.

Client work stays confidential, so these are the projects I can show you in full.

• Eight LLMs graded blind

  One agentic workload over 100,000 business-news articles, run eight times with only the model changed. The most expensive model finished last, and quality peaked mid-range.

  Read the write-up

• Seven adversarial agents

  Claude agents playing Diplomacy through a custom agent-to-agent negotiation protocol, used to measure how agent skill scales with model size and spend.

  See the code

• 1,000 forecasts in one pipeline

  Monthly crime forecasts across Chicago's 50 wards and 20 crime types, R² 0.96 one month out and 0.91 at four months, built end to end with Claude Code.

  See the code

• Fine-tune and deploy an LLM on GPU

  A published walkthrough for fine-tuning an open-source LLM and serving it on GPU in OCI Data Science.

  Read the tutorial

How it goes

Three steps, and you can stop after any of them.

1. 1

   A short call on the piece of work, free

   You describe what is stuck or what is next. I tell you whether I am the right pair of hands for it, and roughly how many hours it should take.

2. 2

   Hands on it, by the hour

   I work in your repository, on your stack, to your conventions, alongside your team rather than instead of it. Hours are reported weekly, and once the scope is clear I can quote a fixed bid instead.

3. 3

   Handed back, documented

   Code, documentation, and a working session, so your team owns what was built and the next piece does not need me.

$250 an hour, no minimum. A retainer from $3,500 a month buys reserved capacity and a standing weekly session. Full rates are on the engagements section of the main page.

Questions

What engineering and data leaders ask first.

• Why not just hire someone?

  Because a requisition takes months and this piece of work needs doing now. Hourly help comes out of a services budget rather than a headcount line, needs no onboarding runway, and stops when the work does. If the work turns out to be permanent, that is a good reason to hire, and I will say so.

• What does it cost, and is there a minimum?

  $250 an hour, with no minimum and no retainer required. Once the scope is clear I can quote a fixed bid instead. If you want reserved capacity and a standing weekly session, a retainer starts at $3,500 a month.

• Who am I actually working with?

  Me, directly. Ph.D. physicist, twelve years building production AI and machine learning, and eight years at Oracle, the last three in its AI Center of Excellence, designing generative AI, agent, RAG, and machine-learning solutions for enterprise clients in manufacturing, energy, finance, public sector, and retail. There is no bench of junior staff behind me.

• What stack do you work in?

  Python and SQL first, with scikit-learn, Spark, vector search, Claude and the Anthropic SDK. My deepest cloud platform is Oracle Cloud Infrastructure, with seven OCI certifications. I work on whatever platform your team already runs, and I will tell you plainly where my experience is thin.

• Do you use AI coding assistants?

  Yes, Claude Code mostly, and it is a large part of why the hours go further. Your policy governs your code: if your team does not allow AI assistants on its repositories, I work without them.

• Can a public agency bring you in?

  Buying rules differ by state, city and agency, so tell me how yours buys and I will fit the engagement to it. I invoice through my own company, JMH Datasciences LLC, and will register as a vendor where an agency requires it.

• Are you tied to a particular vendor?

  No. I am independent: no vendor relationships, no reseller agreements, no referral fees. The recommendation is whatever fits your data and your systems.