I get AI systems from prototype to production.
I help funded teams and local operators turn brittle demos, stalled models, and invisible digital presence into reliable systems people can use.
No fabricated case studies. No hype without measurements.
Diagnose. Define done. Build. Handoff.
Battlefield medicine taught outcome ownership under constraint. AI engineering turned that into systems that are measured, monitored, and reliable.
Same discipline, different buyer.
The startup buyer is buying production AI judgment. The local service buyer is buying a clear path from online invisibility to more calls.
For funded startups
Prototype to Production AI
Founders and CTOs with a model that works in a demo but stalls before launch.
A production-ready AI service with clear acceptance metrics, monitoring hooks, latency and cost controls, and a handoff your team can operate.
See MLOps offer →For service businesses
Local Visibility
Local operators starting in one market, with a model that can scale by niche or region.
Website, Google profile, schema, reviews, and local search fixes that turn searches into calls.
See visibility offer →Operator credibility
Special Forces discipline, applied to shipping AI.
I spent over a decade in U.S. Army Special Forces as an 18D Special Forces medic. That work teaches one thing above everything else: own the outcome.
I brought that to AI. The result is a practical bias toward systems that can be measured, deployed, monitored, and handed off without drama.
Special Forces medic
Outcome ownership, triage, and calm execution when plans meet reality.
M.S. Artificial Intelligence
Graduate AI training across ML, algorithms, NLP, vision, and reinforcement learning.
Production systems focus
The work after the demo: data discipline, evals, monitoring, deployment, cost, and latency.
Remote from Michigan
Senior production judgment without coastal overhead or agency bloat.
See the discipline before you hire it.
A self-contained proof-of-work repo takes a model from notebook to containerized service to edge-optimized artifact, with real benchmark numbers.
| Variant | Mean | P95 | Size |
|---|---|---|---|
| PyTorch eager (fp32) | 0.122 ms | 0.128 ms | 811.2 KB |
| ONNX Runtime (fp32) | 0.067 ms | 0.072 ms | 810.3 KB |
| ONNX Runtime (int8 quantized) | 0.058 ms | 0.060 ms | 221.6 KB |
PyTorch eager (fp32)
- Mean
- 0.122 ms
- P95
- 0.128 ms
- Size
- 811.2 KB
ONNX Runtime (fp32)
- Mean
- 0.067 ms
- P95
- 0.072 ms
- Size
- 810.3 KB
ONNX Runtime (int8 quantized)Best
- Mean
- 0.058 ms
- P95
- 0.060 ms
- Size
- 221.6 KB
What it demonstrates
Notebook to live service.
- Train, export to ONNX, then int8 dynamic quantization.
- 1.81x lower mean latency moving PyTorch eager to ONNX Runtime.
- 3.66x smaller artifact after quantization.
- FastAPI service with health endpoint, prediction endpoint, structured logs, and latency headers.
A lightweight process for serious outcomes.
The goal is not to create a consulting theater. The goal is to find the gap, agree on done, close it, and leave the system operable.
Step 01
Diagnose
Find the gap using evidence, not vague AI or SEO claims.
For AI systems, that means model, data, evals, deployment, monitoring, latency, and cost. For local visibility, that means website, Google profile, schema, reviews, and search-to-call paths.
Projects and writing that show the work.
Public, runnable examples of the technical range behind the consulting practice.
/ featured projects
PlantDoc — Plant Disease Classification
CBAM-augmented ResNet18 for 38 plant disease categories. 97.46% accuracy, 99.21% precision, 99.17% recall on the test set.
- Accuracy
- 97.46%
- Precision
- 99.21%
- Recall
- 99.17%
Real-Time Parking Optimization System
Dynamic pricing, A* routing with traffic, and ML demand prediction across 113 zones in downtown Grand Rapids. Sub-100ms response.
- Response
- <100ms
- Concurrency
- 10,000+
- Zones
- 113
Sepsis Prediction Pipeline
Early-warning ML pipeline for sepsis detection. XGBoost, Random Forest, and Logistic Regression with hyperparameter tuning via Optuna.
- XGBoost AUROC
- 0.9998
- RF AUROC
- 0.9760
- LR AUROC
- 0.8955
HMER — Image to LaTeX Converter
Sequence-to-sequence model that transcribes images of mathematical expressions to LaTeX. CNN/ResNet encoder + LSTM decoder.
- Accuracy
- 62.56%
- BLEU
- 0.1539
From Notebook to Edge — Proof of Work
Self-contained ClelandCo demo: train a small CNN, export to ONNX, int8 dynamic quantization, FastAPI service. 1.81× lower latency, 3.66× smaller artifact.
- Latency
- 1.81×
- Size
- 3.66×
/ writing
From Notebook to Edge: ONNX and int8 Quantization, Measured
A runnable walkthrough of taking a PyTorch model to a small, fast production service: export to ONNX for a 1.81× latency win, quantize the head to int8 for a 3.66× smaller artifact, and measure it honestly.
Why AI Prototypes Stall Before Production
The model is rarely the whole problem. The gap is usually data discipline, evals, deployment, cost, and monitoring — and that gap is what most projects underestimate.
The Local Visibility Checklist: 7 Gaps Costing You Calls
Most local service businesses don't have an SEO problem — they have a handful of specific, fixable visibility gaps. Here are the seven I find most often, and how to close them.
Serious work, plainly scoped.
The brand should feel technical and grounded because the work is technical and grounded.
What I will not do
- No fabricated outcomes or fake case studies.
- No mass-spam outreach tooling.
- No AI claims without measurable proof.
- No vague retainers without scope and expectations.
What you can expect
- Evidence before opinion.
- A written definition of done before build work.
- Documentation and handoff as part of the deliverable.
- Pricing tied to outcomes, not theater.
Start with the gap
Have a model stuck before production, or a business customers cannot find?
Tell me where the system is breaking. I will tell you straight whether I can help, what the first paid step should be, and what not to waste money on.