Selected AI builds
Systems I've designed and shipped.
Projects I lead or built — each takes something from the AI frontier (fine-tuning, RAG, OCR, LLM evaluation) and turns it into a tool a classroom can actually use.
Flagship · founder & builderCohort 1 delivered · July 2026
TheAIProgram.org · AI for High School Teachers
A complete learning platform I built from scratch — live at TheAIProgram.org. Its flagship, Iowa's first statewide AI course for teachers, delivered its first cohort in July 2026 — and a Quantum ML & AI course (QAI 1010) is next.
- Full LMS: enrollment, auto-graded exercises, a Canvas-exportable gradebook, discussions, and course certificates.
- In-browser Python (Pyodide) with CodeLens line-by-line execution — no install, no setup, built for beginners.
- A Socratic AI tutor on every step: a three-rung hint ladder (nudge → strategy → worked step) that deliberately withholds full solutions.
- Anchored to the UNESCO AI Competency Framework — teachers from any subject, statewide, no coding required.
TheAIProgram.orgPyodideSocratic AI tutorFull LMSUNESCO framework
LLM-as-judge · eval
AI Arena
An automated arena that grades student LLM, RAG, and GenAI projects — the open-ended assignments Kaggle InClass can't score.
- LLM-as-judge scoring with a primary evaluator and cross-family re-grades, so results don't inherit one model's bias.
- Live leaderboard and submission flow on FastAPI + Cloud Run, Firebase Hosting, and Firestore.
- Built to be reused across AI 2010, data-science courses, and Iowa teacher bootcamps.
FastAPICloud RunFirestoreClaudeGemini
Project lead · LLM fine-tuning
PyTutor AI
I lead PyTutor AI — an open-weight LLM I fine-tuned for Socratic teaching, guiding questions first — supported by ISU's Office of the Provost and the Department of Computer Science.
- Custom Socratic fine-tune of a 27B open-weight model for CS, DS, ML, and AI coursework — SFT + DPO with a pedagogical reward model that favors hint-first responses.
- Five-benchmark evaluation stack that scores teaching quality, not just answer accuracy.
- Trained and served on Nova, ISU's GPU research cluster.
27B fine-tuneSFTDPOReward modelsNova GPU
Lead technical architect · since 2010
LIRT: Retrieval at National Scale
I designed and lead the technical architecture of the Legislative Institutional Repository of Thailand (LIRT), the country's largest institutional repository by document count: about 120,000 mostly Thai-language legislative documents.
- RAG search over the full collection (OAI-PMH ingestion → vector index), so legislative records can be asked about in plain language.
- OCR ingestion that turns scanned PDFs and archives into retrievable, citable knowledge.
- Embedding and inference jobs scaled across GPU nodes on ISU's Nova cluster; Qdrant, pgvector and Gemini File Search evaluated head to head. Contributor to the NIST TREC RAG track.
LIRTRAGOCRQdrantNova GPU
Benchmark · NIST TREC 2026
Video Question Answering
Team iastate's entry in the TREC 2026 Video QA track tested how far off-the-shelf multimodal models go with no fine-tuning, when runs are chosen by calibration rather than by hindsight.
- Vote and rank-fusion ensembles over off-the-shelf multimodal models; no fine-tuning.
- Runs selected on 573 calibration questions rebuilt from the 2025 data, with expected results recorded before any 2026 results came back.
- Our best multiple-choice run had the highest Top-1 accuracy of 42 submitted runs, within a statistically tied top group. Talk at TREC 2026, November 18.
Multimodal LLMsRank fusionEvaluationNIST TREC
Multi-model · LMS
Multi-Vendor AI Teaching Assistant
A course AI assistant that runs on several models at once — so students aren't locked to one vendor, and neither is the university.
- Deploys open models (Llama 4 Scout, Phi-4, gpt-oss, Gemma, Nemotron) on Azure AI Foundry for ~100 students per course.
- A Canvas-LTI RAG "living textbook" grounded in each course's own materials and owned by ISU.
- Multi-vendor by design (Claude / GPT / Gemini) to avoid single-vendor lock-in.
Azure AI FoundryCanvas LTIRAGOpen models
LMS automation · auto-grading
Canvas Course Copilot
Python tooling that auto-generates course content and quizzes, then wires up automated grading inside the LMS.
- Generates Canvas modules and QTI 1.2 quizzes straight from lecture slides, using Python + lxml.
- Gradescope ↔ Canvas grade-sync workflows that hold up at large-course scale.
- Cuts the manual work of authoring and grading hundreds of assessments.
CanvasQTI 1.2PythonlxmlGradescope
Workshop · enablement
Build Your Own AI Assistant
A hands-on workshop series teaching educators and students to stand up their own RAG assistants — local open-weight models vs. hosted APIs.
- Walks through the full local-vs-cloud tradeoff: cost, privacy, latency, and quality.
- Includes automation patterns with n8n and KNIME for data and grading pipelines.
- Delivered through ISU's Student Innovation Center and the D4 bootcamp.
RAGLocal inferenceAPIsn8nKNIME
