Healthcare AI, fraud, audio ML, agents, humanitarian analytics. The common thread is caring why a model works, not just whether it does.
A local League of Legends matchup assistant that auto-detects your champ-select pick and answers from a precomputed lookup table instead of generating live.
Under 200ms6,512 same-lane pairs, tiered precompute, packaged as a Tauri desktop app
Responsible AI frameworks govern the model, not the land, water, and power grids it runs on. A case study of the Digital Gateway data center fight in Prince William County.
$750M/yearstate tax subsidy cost, concentrated in a handful of localities
Mapping Anthropic's and Google's public safety reporting onto NIST's AI Risk Management Framework — and showing why even the best voluntary compliance still leaves the company in control of what gets disclosed.
4 functionsGOVERN, MAP, MEASURE, MANAGE compared across two frontier labs
A LoRA fine-tune of Qwen3-8B that generates a single MCP tool call — or abstains — instead of hallucinating tool names and malformed arguments.
75% call accuracyup from 56% pretrain; abstention accuracy more than doubled
An MCP server that reconciles real LLM inference cost from local usage logs across Claude Code, Codex, OpenCode, and Amp.
Multi-toolnormalizes ccusage output across four CLI agents into one cost model
Point it at YouTube videos and it transcribes, segments by topic, summarizes two ways, extracts entities, and graphs what recurs across videos — no paid APIs.
Zero paid APIsfaster-whisper + local Ollama, entities graphed in Neo4j
A Vision Transformer for pneumonia detection with a Grad-CAM overlay showing exactly which pixels drove the call.
Under 10sFastAPI + ViT inference, React/TypeScript front end
An LLM that turns raw emergency-department data into structured clinical risk summaries.
118,385 encountersStanford MC-MED dataset
Does scrubbing noise before the model sees it actually help keyword spotting? Sometimes — and it depends on the architecture.
94.76% clean accuracya CNN that beat the transformer in every condition
An Oregon Trail-themed step tracker: your steps move a wagon, and missing your goal means dying of dysentery.
One weekend, four peopleHoosHack 2026, live on a physical iPhone
Tell it what you want to learn and what you already know; it returns a prerequisite-aware roadmap and schedule.
2,759 courses3,682 prerequisite links, local Llama 3.1
A head-to-head on NYC taxi data to find where the library choice actually pays off, and where it doesn't.
15× faster loads0.0032s vs 0.0463s, 100 runs each
Portfolio performance, attribution, and risk against historical market data. Python pulls and processes; Tableau shows.
Built in phasesvaluation first, risk analytics on top
A pipeline that finds fraud across 6.3M+ transactions where the honest answer is almost always "not fraud," now deployed live on Azure.
94% recall0.99 ROC-AUC, at a 0.13% fraud rate
Detecting temporary shelters in aerial imagery so aid groups could reach displaced people faster.
99% recallblue-tarp detection, six models compared
Predicting hourly electricity demand for American Electric Power from 14 years of history.
1,644 MW RMSE~121k hourly observations, 2004–2018
How length of stay maps to mental health for international students in Japan, from 268 survey responses.
Peaks in year 3PHQ-9 of 9.09, not year 1 as you'd expect
A pip-installable Python package for Monte Carlo simulations with customizable, weighted dice.
3 core classesDie, Game, Analyzer — pip-installable from GitHub
A group project testing Blue Nile's own claims about what drives diamond price and appearance across 1,214 diamonds from their stock.
89.6% of varianceexplained by carat alone, across 1,214 diamonds
A group project using logistic regression on Pew Research survey data to test how internet access in Singapore relates to social media use and opinions of the US and China.
1,000 respondentsPew 2022 Global Attitudes Survey, SAS logistic regression