- Built an open-source Python tool that plants controlled failures in passing agent traces, reruns a team’s DeepEval metrics, and fails CI when those evals miss wrong tool arguments, repeated side effects, hidden tool errors, or damaged responses.
- Published a versioned golden set with 24 cases, 13 tool contracts, and 263 pinned mutations across ten domains. A response-only evaluator missed 176 faults while a complete trace-and-outcome evaluator caught all 263.
Daniel Gaskins
Applied Machine Learning Engineer
hello@danielgaskins.com danielgaskins.com github.com/danielgaskins linkedin.com/in/daniel-gaskins-ml United States · Open to relocation
Summary
Applied AI engineer and founder who has built document-processing, computer-vision, agent-evaluation, and business-automation systems. Strongest at turning an unclear problem into a testable workflow, finding where it fails, and carrying the result into production.
Technical Skills
Machine Learning: Agent evaluation, golden datasets, LLM document extraction, model training and evaluation, experiment design, error analysis, data leakage, train/serve parity, NLP, computer vision, gradient-boosted trees, CLIP
Frameworks: PyTorch, TensorFlow, scikit-learn, LightGBM
Software & Systems: Python, C++, TypeScript, JavaScript, Node.js, React/Vite, REST APIs, OAuth 2.0, GCP/Firebase, Cloud Functions, Firestore, Cloud Storage, testing, CI/CD
Selected Technical Projects
- Built a code generator that turns trained, one-output LightGBM models into dependency-free Python, C++17, or JavaScript raw-score inference.
- Added executed and compiled parity tests across all three runtimes at 1e-12 tolerances, including missing-value routing and explicit rejection of unsupported model types.
Professional Experience
SyncABill — Founder & Applied AI Engineer
2025–Present
- Built and deployed a React/Vite invoice application with TypeScript and Node.js services on Google Cloud/Firebase, using Firestore for workflow state and Cloud Storage for source documents.
- Connected structured document extraction to deterministic checks, controller-configured review rules, and QuickBooks Online or Xero.
- Completed Google’s CASA Tier 2 security assessment and brought the product to SOC 2 readiness.
Casabauhaus — Founder & Operator
2022–Present
- Operate a vintage-furniture business spanning sourcing, pricing, merchandising, sales, and delivery, supported by agent and API workflows for reporting, social media, seasonality analysis, and Shopify.
- Developed furniture-valuation and mid-century classification models used in sourcing decisions, including a fine-tuned CLIP model built before modern vision-language models were widely practical.
Perfsy — Founder & Machine Learning Engineer
2021–2022
- Built a computer-vision pipeline to extract structured fields from scanned DMV vehicle titles.
- Normalized scan orientation with OpenCV and experimented with character primitives, autoencoders, convolutional features, and DNN classifiers.
- Created data and evaluation workflows to inspect recognition failures and improve the pipeline.
Education
University of California, San Diego
Condensed Matter Physics | 2013–2016