APPLIED AI R&D
Economic agents, reusable numerical models, and embedded computation. Our research connects explicit plans to specialized tools and keeps the evidence behind each capability visible.
RESEARCH PORTFOLIO
Explore the prototypes, their current evidence, and the work still ahead.
A working research system that turns economic and forecasting questions into typed computational graphs, with reviewable plans, tool calls, and output artifacts.
Explore this program →Reusable numerical packages for statistics, factor models, simulation, and conditional distributions, connected to Rust research workflows through explicit data and library interfaces.
Explore this program →An early computational workbench for drone-related numerical kernels, starting with Rust reference calculations and an ARM assembly implementation checked under QEMU.
Explore this program →Research into bounded agents that help service businesses plan work, allocate operating resources, and evaluate decisions without replacing accountable operators.
Explore this program →INSIDE THE PROJECTS
Agent Factors BM combines a Rust graph planner and execution engine with Model Context Protocol (MCP) tools. Forecasting components support panel preparation, factor fitting, innovation models, simulation, diagnostics, and ensembles. Companion UI and server projects provide workflow sessions, authentication, and persisted runs.
A documented synthetic trace demonstrates a real MCP client/server handshake, tool discovery, and a target grid for two assets and two horizons. Focused tests cover graph validation, argument propagation, and route preservation. These establish parts of the execution system, rather than forecast accuracy or investment performance.
Data provenance, model assumptions, and human authorization remain explicit. Research outputs do not authorize trades.
Public console unavailable
The Mojo packages include moments and covariance estimators, principal-component analysis (PCA), ordinary least squares and vector autoregression (OLS/VAR), stochastic simulations, and conditional normalizing flows. Shared model contracts and C-compatible library boundaries let Rust graph tools reuse these numerical routines.
PCA has a deterministic CPU reference; its GPU model currently reports a CPU fallback. Conditional spline flows provide CPU reference inference and differentiation, while their graph-facing library interface exposes inference. Training-only fitting and frozen held-out transforms are explicit parts of the factor-model workflow.
Backend status, training intervals, buffer ownership, and artifact provenance stay visible. GPU execution is described only where implemented.
Public console unavailable
The X500 PX4 Sim workbench is development around the Holybro X500 V2 platform. Its current implemented scope is a three-dimensional vector dot product: Rust calls an ARMv7-A assembly routine through a C ABI, with a Rust reference calculation for comparison.
Linux ARM binaries run under QEMU to check instruction and program logic. Host builds exercise the Rust fallback. Neither path reproduces aircraft sensors, flight dynamics, real-time timing, or Pixhawk firmware behavior.
The current workbench sends no flight commands and accesses no aircraft hardware. Hardware integration requires a separate target and validation effort.
Public console unavailable
OPERATIONS RESEARCH
Research into bounded agents that help service businesses plan work, allocate operating resources, and evaluate decisions without replacing accountable operators.
Operator approval remains required for customer commitments, purchases, scheduling changes, and other external actions.
Public console unavailable
R&D COLLABORATION
We welcome research partners who can define the operating question, relevant evidence, and what a useful result would look like.