Think With Me
Each bubble is a rabbit hole. Click one and fall in.
BMW asked: which cars will actually sell?
Built ML models on real dealer data to predict which vehicle specs move fast. LightGBM + Optuna-tuned TabularMLP, served through a live API so dealers make spec decisions without needing a data team.
what happens to the reward when Alexa won't shut up?
Trained an RL policy that learns WHEN to intervene in multi-agent dialogue — not just what to say. Reduced unnecessary interruptions 25% while keeping task success intact. Currently writing this up for NeurIPS.
finding humans in a blizzard, from 200 feet up
Fine-tuned Faster R-CNN on thermal imagery with snow, smoke, and sensor noise augmentation. 20% recall improvement in adverse conditions — the kind of gain that means someone gets found.
a therapist that remembers everything but never judges
RAG over 8 psychological frameworks (CBT, IFS, NVC) to pattern-match emotional entries. Tracks cognitive distortions, mood trends, and recurring triggers across sessions.
the math proving Durham's crosswalks are placed by bias
Black residents: 32% of Durham's population, 47% of pedestrian crash victims. Built interactive maps comparing AI allocation vs. need-based infrastructure using Census + NCDOT crash data.
shazam, but for outfits in movies
FashionCLIP embeddings + FAISS vector search over 20K garment crops. Four-stage recommendation pipeline (FAISS → NeuMF → SASRec → diversity filter) that actually surfaces style you'd wear.
shipping containers predict your grocery bill
SARIMAX pipeline combining port traffic volume with CPI data. 0.67–1.69% MAPE across major categories — turns out boats know inflation before the Fed does.
how much carbon did that prompt just cost you?
Chrome extension estimating energy, carbon, and water footprint per AI prompt in real-time. Privacy-first, client-side only, with a daily impact dashboard.
semantic search that reads case law like a lawyer
Embedding-based retrieval with citation-link modeling to surface related precedents. Built for explainable legal research — not just keyword matching, actual meaning.
can your model pass the EU AI Act?
TF-IDF + Logistic Regression for risk classification, rule-based article evaluation (Articles 5, 6, 9, 10, 14), automated remediation. Because compliance shouldn't need a lawyer and a data scientist in the same room.