Projects
Technical Work
Quorum: Investment Consensus
RL-based portfolio management with cross-asset attention and regime detection. Data in, decisions out. End-to-end reinforcement learning.
Cadence: Orchestrating Life
One action propagates through a knowledge graph. Frequencies evolve by survival of the chosen. A system that learns what you do. And adapts.
Resonance
Seven APIs feed a unified knowledge graph. Frequencies evolve from listening behavior. The system learns your taste and finds what you're missing.
Primordial
Organisms evolve bodies and brains from scratch. Spring-mass physics. Neural network locomotion. No gradient descent, just survival. A multi-part investigation into emergent complexity.
Noloss
Creatures evolve on a tiny 3D planet. Attention-based brains, context-sensitive actions, and a food web that produces herbivores, predators, and social species from a single architecture. You shape the world. Evolution shapes the creatures.
fold
An agent folds its own transcript instead of losing it. Stale spans collapse into briefs while the exact originals stay on disk, verified by hash. Part 1 measures the bill: the same work as native compaction at 0.60x the tokens. Part 2 measures the memory: values that live only in the transcript survive folding 60 of 60, while the compacting arm loses the one key it needs to continue.
canon
Project memory for a coding agent: one article per asset at a knowable address, an append-only journal beneath it, and recall that arrives on its own when a tool call touches the asset. Part 1 measures it whole: 19 of 20 traps avoided against 8 of 20 for a no-memory floor. Part 2 prices touch and recommendation directly, across six preregistered studies and 679 graded cells.
shedding-bench
A working-memory stress benchmark for coding agents: what a session still knows after its context has been shed, and what getting it back costs. The planted material is the structure of the work rather than a fact to remember, collection is a stale file to fix with wrong values already filled in, and grading joins what the agent wrote to hashes of what was delivered. Any harness can be the subject.