Index → Projects

Work
in motion.

A selection of projects, notebooks, and experiments. Some shipped, some in motion, some still ahead. Each one is a different angle on the same study of intelligence.

An avatar and recommendation system for sake recommendations and food pairing. The challenge is not simply recommending a bottle, but understanding why a person likes what they like — combining structured tasting data, flavor embeddings, and food-pairing knowledge with conversational interaction. The long-term vision: a digital avatar that acts as your personal sommelier.

Recommendation systems Preference modeling Embedding spaces Knowledge graphs Food-pairing intelligence Conversational AI Digital avatars

A production forecasting platform running live on GCP, predicting daily behavior for 500 content series at horizons up to 180 days. Real-world forecasting is never one model: the system backtests a multi-family model zoo (statistical, gradient-boosted, deep, and in-warehouse BigQuery ML) and automatically selects a champion per series, per horizon. New series with no history are forecast from day one using a digital-twin method: launch curves from similar series, weighted by a hype signal, refined by pattern clustering once real data arrives.

Every forecast is written to an append-only ledger and scored against actuals as they land, so accuracy at the end of the horizon is measured, not simulated. Fully reproducible from a single repo: Terraform-managed infrastructure, keyless CI/CD, monitoring, alerting, and a teardown path.

Long-horizon forecasting Cold-start digital twins Multi-model champion selection Probabilistic intervals Production MLOps on GCP Infrastructure as code

A real-time computer vision pipeline that reads facial expressions from a webcam, builds a representation of what the face is doing, and drives a generative avatar to mirror it back as the user reacts.

The interesting questions sit between three problems: what does a face actually communicate, how do you encode that into a representation a model can manipulate, and how do you render a believable mirror without uncanny lag.

Sensing

The eye that watches.

Real-time facial landmark detection, expression vectors, and gaze tracking from a single webcam feed.

Modeling

Human to twin.

Mapping landmark geometry into a pose-and-expression space the avatar can act on. The translation layer between human and twin.

Mirror

Where uncanny becomes useful.

Generative avatar rendering that follows the face closely enough to feel like reflection.

A study at the intersection of computational neuroscience and AI — what the brain teaches us about agents that perceive, predict, and adapt. The current focus is active inference: the idea that cognition is, at its core, a process of prediction.

The predictive brain.

The free-energy principle frames cognition as a continuous loop of prediction-minimization. A brain — or any agent — maintains a generative model of the world and acts to shrink the gap between what it expected and what it observed. The connection to modern self-supervised learning is direct: prediction error is the training signal. The connection to organizational decision-making is what makes the theory interesting outside the lab: teams, like brains, are systems that update beliefs from feedback, and the ones that update faster usually win.

  • Free-energy principle (Friston) — formalism behind the loop.
  • Predictive coding — perception as inference, not detection.
  • World models — planning inside a learned generative model.
Open Source

Live on GitHub.

Notebooks, experiments, and works in progress land on GitHub as they happen.