Builds
The engineering layer of the site: engineering practice, deeper systems work, and shorter prototypes brought together so the reader can see both the long-build work and the faster experimental edge in one place.
Engineering Practice
FAIRE — Frontiers in AI Research and Engineering
An agentic frontier AI wiki that turns scattered learning into directed building: 10 canonical tracks, concept pages, primary-source references, and Minimum Valuable Builds that push each topic toward implementation.
Engineering Practice Log
A running place for the engineering I want to compound for the future of AI, ML, and data systems: evaluation, agents, tooling, data systems, inference, reliability, and end-to-end system design.
Deep Builds
No-Exemplar Continual Learning via Causal Invariance
Neural networks forget old tasks when learning new ones. Instead of storing past examples, this asks why forgetting happens — and finds the answer in causal structure. Outperforms replay-based baselines with zero stored images.
HMV-CRL: Separating Platform Influence from Genuine Preference
Recommendation engagement conflates what users want with what the algorithm pushed. This model learns both representations separately — and finds that platforms amplify engagement while suppressing the genuine preference signal.
Zero-Shot Coordination in Multi-Agent RL
Agents trained separately fail to coordinate when paired with a stranger — they built different habits. This asks what training structure gives agents conventions general enough to work with anyone.
Causal State-Space Models for Time Series
Forecasting models predict what will happen — but decisions change the pattern. This integrates causal structure into sequence models so they can reason about interventions, not just correlations.
Port-Hamiltonian Neural Networks
Neural networks predict physical motion but don't know energy is conserved — over long predictions, they drift. This bakes conservation laws into the architecture so violations are impossible.
Generative Social Network Simulation
Social simulations use hand-coded rules. This uses language models as the agents themselves — each with distinct beliefs and tendencies — to study how opinions spread and polarize across a network.
Personality-Aligned Vision-Language Model
People describe the same image differently depending on how they think. This aligns a vision-language model to specific personality styles, then tests whether that consistency makes it a better coaching assistant.