Applied AI / machine learning engineering
I’m a founder and applied ML engineer. I turn ambiguous questions into testable experiments, inspect where models fail, and build useful systems around what survives.
01 / SELECTED WORK
Recent work starts with an uncomfortable question: how could this result be wrong? I turn that question into a test, build the surrounding system, and keep the failure visible long enough to learn from it.
Gemini can read an invoice, but it should not decide what enters a company’s books. I built SyncABill to extract the invoice, check the result, apply the controller’s review rules, and send approved bills to QuickBooks Online or Xero.
Your agent tests may all pass and still miss a broken tool call. Mendmark plants controlled faults in passing traces and includes a public golden set showing how response-only, trace-only, and complete evaluators behave across 263 mutations.
A trained, one-output LightGBM model becomes dependency-free Python, C++17, or JavaScript. Executed and compiled tests compare every target against LightGBM raw scores at 1e-12 tolerances.
At Perfsy, I built a computer-vision pipeline to extract structured fields from scanned DMV vehicle titles. I normalized scans with OpenCV and iterated across character primitives, autoencoders, convolutional features, and DNN classifiers.
I operate this vintage-furniture business and use agent-assisted, API-driven workflows for pricing, reporting, seasonality analysis, social media, and Shopify. Earlier models estimated furniture value and classified mid-century pieces, including a fine-tuned CLIP model built before modern VLMs became practical.
02 / HOW I WORK
model.fit().The dataset is messy. The metric lies. Latency enters the room. I like that part: finding the failure, tightening the loop, and turning a promising model into software people can rely on.
Start with a question. Build the dataset, run the experiment, and keep honest notes.
Write the test that could prove the idea wrong. Read the misses, not just the mean.
Give the model dependable data, inference, monitoring, and deployment paths.
Put it in front of a real user and stay close enough to see what breaks.
METHODS + TOOLS
03 / BLOG
Short technical notes about agent evaluation, model failures, and the systems built around uncertain outputs.
Every review consumes a person's attention. A practical way to decide which AI actions should stop, what the reviewer needs to see, and what the system must never decide alone.
Read the field notePlant one known failure, rerun the same evals, and see whether the score changes. A practical introduction to mutation testing for agent tool use.
Read the field note04 / ABOUT
I studied matter at its smallest scales. Then I started companies. Both taught me the same thing: a good answer begins with a better question.
Daniel Gaskins is a founder and applied AI and machine learning engineer who builds reliable systems around models, tools, and real business processes. He created Mendmark, a mutation-testing system for agent evaluations; built lgbm-to-code for dependency-free cross-runtime model inference; and founded SyncABill, an AI-assisted invoice workflow.
Daniel studied condensed matter physics at the University of California, San Diego before moving into founder-led software and machine-learning work. His experience spans agent evaluation, document AI, computer vision, production workflows, and model deployment. He works from failure cases outward: define what must remain true, build the smallest useful system, and test whether it holds up outside the demo.
05 / CONTACT
If you are training or evaluating models, building the systems around them, or turning research into a product, I would like to hear what is hard.