I help agent and RAG teams evaluate, understand, and optimize the systems they are building—especially tool calling, skills, retrieval, routing, and model choices. I bring machine learning discipline to agent development so you can find the biggest lift, protect the weakest link, and scale on measured results.
Agent development is moving at software speed inside a non-deterministic, statistical world. That creates enormous opportunity—and dangerous blind spots.
I work across diagnosis, experimental design, evaluation infrastructure, optimization, implementation, and technical education. The goal is practical: understand what is happening, identify leverage, and make the system materially better.
Create modular, testable evaluation criteria grounded in how users and domain experts actually judge outcomes.
Reduce ambiguity in tools, MCPs, and skills. Test naming, descriptions, grouping, masking, and exposure strategies.
Match jobs to models, prompts, tools, skills, and retrieval strategies using structured experiments instead of noisy averages.
Diagnose retrieval failures and shape growing knowledge into navigable, task-specific context rather than a flat pile of chunks.
Analyze request distributions, tool calls, transitions, outputs, failures, and changing usage patterns over time.
Hands-on workshops and working sessions that make experimental design, uncertainty, classification, ranking, and optimization directly useful to agent developers.
AI is growing like a forest: fast, fertile, tangled, and full of opportunity. I help turn that growth into an orchard—structured enough to inspect and harvest, dynamic enough to keep adapting.
Identify promising patterns, successful skills, reliable routes, and high-value use cases. Give them the evidence and structure needed to scale.
Retire weak configurations, redundant tools, misleading descriptions, noisy context, and branches that create cost without creating value.
Deploy winners, improve routing, codify learning, and bring tools, skills, evals, and decisions back into a dependable operating system.
I can enter at the point where a system is unclear, unreliable, expensive, difficult to scale, or simply not living up to its promise.
Map the architecture, traces, decisions, metrics, and user expectations. Find the places where uncertainty is hiding.
Create the smallest credible evaluation or experiment that can separate signal from noise and reveal leverage.
Build the evaluation pipeline, analysis, routing logic, data structure, dashboard, or workflow needed to make the insight operational.
Integrate the result into LangSmith or your existing stack, train the team, and establish a loop that keeps learning from production.
Knowledge Algorithms grew out of Energy Algorithms, and from the work of helping build one of the Midwest's most successful virtual electricity trading companies. The operating belief was simple: wild and chaotic data should be cultivated into enough structure to be accessible at any moment, while remaining dynamic enough to adapt to a changing market.
Electricity markets can move in extreme and unexpected directions. That volatility is not only risk—it is the source of opportunity. The work demanded quantitative discipline, fast feedback, strong judgment, and systems that could evolve without becoming opaque.
Agents and RAG systems are proliferating quickly. Tools, skills, prompts, retrievers, models, and orchestration layers are growing in every direction. The creativity is real. So is the chaos.
I do not want to reproduce the same old software patterns with an LLM attached. I want to work where the problems are difficult and the upside is massive: evaluation, context, tool use, routing, optimization, knowledge structure, and continual learning.
My working rhythm is direct: assess risks, grab opportunities, iterate. I analyze pain points, verify operation, research alternatives, design experiments, implement solutions, and communicate clearly with technical and non-technical teams.
The infrastructure must be ready to pivot, identify the new thing, and remain robust under deadline. Data must be labeled, modeled, and structured correctly. Teams need visual observability—enough to know where they are without distraction, right now. A solid core is what makes creative, dynamic work possible.
I am especially interested in working with LangChain and LangSmith teams and customers, where small evaluation and orchestration ideas can become scalable infrastructure. The outcome I aim for is confidence that insight can be reached from every part of the data stack, whatever the system encounters.
These are not rigid product packages. They are recurring solution patterns that can be adapted to a company's data, architecture, and operating needs.
Gather skill versions, group and contrast them, evaluate them against real work, and rewrite them for the use case. Preserve the creativity and analytical power of the company in an evolving, testable space.
Create a dynamic hierarchy over tools, MCPs, and workflows. Reduce context burden, expose clearer contrasts, control gateways, and track the paths agents follow.
Turn documents, tools, skills, and internal knowledge into hierarchical, visual structures that both people and agents can traverse without flattening everything into a ranked list.
My projects explore the same underlying question from different angles: how can data and decisions be structured so that people and agents can act with less friction and more confidence?
My practical harness-optimization method for separating model, prompt, and tool effects using best-worst evaluation instead of noisy averages.
Read on TDS →Zero-shot embedding classification using cross-modal pairwise structure; improved text/image agreement from 61% to 89% in the published experiment.
Read on TDS →A research metric connecting Fréchet- and KL-like distributional differences for feature selection and interpretable embedding analysis.
Read paper →A browser research vault combining capture, semantic structure, topic maps, and retrieval for a more useful personal knowledge system.
Open project →Hierarchical visual navigation and alternative controls for content-heavy and accessibility-sensitive experiences.
Ask for a demo →Essays and working ideas on machine learning, decision systems, agent evaluation, and useful structures for complex problems.
Read the Substack →You do not need less ambition. You need a clearer view of what is working, what is not, and where the next unit of effort creates the greatest lift.