Research & Publications

Exploring topics through academic research and investigation

Featured I Fine-Tuned an 8B Model to Stop Being Agreeable. It Runs on a Single CPU for $0/Month

I Fine-Tuned an 8B Model to Stop Being Agreeable. It Runs on a Single CPU for $0/Month

May 2026

I Fine-Tuned an 8B Model to Stop Being Agreeable. It Runs on a Single CPU for $0/Month. Carole is a chatbot built for neurodivergent users — specifically those with rejection-sensitive dysphoria. She validates first, then redirects. Not “great idea, here’s why you’re wrong.” Actual validation. Then a question that reframes the problem. The whole stack — fine-tune, RAG, inference, gated web demo — cost under $20 to build and nothing to run.

Read More View HuggingFace
Spatial Signal Transfers, Temporal Signal Does Not: A Real-Data Re-evaluation of Population-Level Privacy Decay

Spatial Signal Transfers, Temporal Signal Does Not: A Real-Data Re-evaluation of Population-Level Privacy Decay

May 2026

A widely-cited synthetic benchmark for “temporal privacy decay” reports test R² = 0.998 — a value that, taken at face value, suggests the rate at which personal data ages out of sensitivity is sharply learnable from a population of users. We re-run that benchmark on three real datasets — GDPR fines (n = 212), HIPAA breaches (n = 1,632), and Microsoft GeoLife GPS trajectories (n = 48,406 records

Read More View
Attuned Resonance: A Multi-Model Cascade for Inbound Call-Center Routing

Attuned Resonance: A Multi-Model Cascade for Inbound Call-Center Routing

May 2026

We present Attuned Resonance, a cascade of four specialized models for pre- dicting and optimizing inbound call-center routing. Given a caller transcript and (optionally) caller audio, the system predicts: (1) call intent, sentiment, urgency, complexity, and Jung x Campbell archetype features (Intake NLP); (2) caller voice tone across 10 emotion classes (Voice Tone Classifier); (3) expected han- dle time, first-call resolution probability, and CSAT for each candidate advisor (Outcome Predictor); and (4) the optimal advisor assignment that maximizes CSAT while respecting burnout constraints (RL Router).

Read More View HuggingFace
Predicting Scripted Outcomes: Lessons from Building an ML System on 482K Pro Wrestling Matches

Predicting Scripted Outcomes: Lessons from Building an ML System on 482K Pro Wrestling Matches

Machine Learning May 2026

We describe Ringside Analytics, an end-to-end machine-learning system trained on 40+ years of professional wrestling data — 482,166 matches and 731,133 wrestler-match par- ticipations spanning WWE, AEW, WCW, ECW, NXT, and TNA.

Read More
CANO: Context-Aware Noise Optimization for Adversarial Privacy Protection

CANO: Context-Aware Noise Optimization for Adversarial Privacy Protection

May 2026

We additionally train a Deep Q-Network (DQN) policy through adversarial co-evolutionary training, where the defense adapts against an attacker that continuously retrains on protected data. We evaluate CANO against five baseline strategies (Gaussian, FGSM, PGD, Carlini-Wagner, and Laplace noise) across 12 datasets (including the real FP-Stalker browser-fingerprint corpus — 776 users, 13,674 fingerprints), 3 attack models, and 6 noise budgets, totaling 68,885 experimental configurations (54,281 in-scope after excluding the 2-user cybersec_intrusion dataset).

Read More View HuggingFace Kaggle
AI Infrastructure Economics: A Complete Decision Framework for Data Center Location, Hardware Selection, and Climate Risk — 2025–2026 Research Series

AI Infrastructure Economics: A Complete Decision Framework for Data Center Location, Hardware Selection, and Climate Risk — 2025–2026 Research Series

Data Centers Apr 2026

A four-part research series quantifying the full cost of AI data center deployment — from global site selection to hardware lifecycle to device-level TCO. Using Monte Carlo simulation, machine learning, and real government datasets, the research demonstrates that Nordic locations are 75% cheaper than traditional US hubs, hardware energy costs dwarf facility costs by 5–10×, and climate change compounds that gap every year through 2050. Includes full Python source code, reproducible notebooks, and supporting cost models for hardware tiers and Apple ARM devices.

Read More View Kaggle
Financial-Structural Vulnerability: How Private Equity Ownership Architecture Produces Consumer Harm in Essential Services

Financial-Structural Vulnerability: How Private Equity Ownership Architecture Produces Consumer Harm in Essential Services

Consumer Protection Apr 2026

Private equity ownership of essential consumer services produces measurable harm when aggressive leverage meets weak federal oversight.

Consumer Protection Marketing Finance
Read More View
Featured We Trained Neural Networks to Predict Where the ISS Will Be in 6 Hours

We Trained Neural Networks to Predict Where the ISS Will Be in 6 Hours

Orbital Mechanics Mar 2026

TL;DR: Our LSTM predicts ISS position to within 125 km at 6 hours (54.5 km at 1 hour) — 10x better than physics-based propagation. Adding solar wind data via cross-attention improves predictions by 17% during geomagnetic storms.

Machine Learning
Read More View