Modeling the Physical World

Andrew Diaz

Computer Vision + Prediction Infrastructure for the physical world.

Andrew Diaz
"Trying to listen."

Computer Vision + Prediction Infrastructure for the physical world.

Andrew Diaz

"Do the work."

The Thesis

The physical world generates data that nobody is reading. A tennis player's split-step latency. A customer's stall time before ordering at a brick-and-mortar counter. A retail food operation's demand rhythm. A warehouse worker's route deviation. These signals are ambient, continuous, and invisible to most models trained on public datasets.

I studied electrical engineering — signal processing, the math of systems that sense and respond. A graduate course in computer system design taught me that every complex system is just a file system — structured layers of abstraction all the way down. That principle scales: from operating systems to prediction pipelines to the physical world itself.

The same architecture extends everywhere. Sports is the proof of concept — a fast, measurable odds market with real capital at stake. If the pipeline works there, the machine learning is sound, and it works anywhere physical-world behavior generates signal. Restaurants. Retail. Logistics. Every venue is an unmonetized sensor network.

A deployed sports prediction model put $20K of my own capital at risk and is up $4,304 (+21%) — purely statistical, CatBoost on historical match data, no computer vision. Autonomous agent systems I've built — multi-agent workflows orchestrating parallel models, content engines, and decision pipelines — prove that one person with agent orchestration can now operate at the pace of a full engineering team. Different domains, same discipline: capture the signal nobody else is reading, automate the response, ship it.

The edge isn't better models. It's better data. Build the sensor.

EXPLOSIVE: 0 m/s²
REACT: 0ms
LATERAL: 0 m/s
PHASE: 0
REACT
+0ms
READY
SPLIT
PREPARE
TRACK
LOAD
STRIKE
FOLLOW
Andrew Diaz
Education
B.S. Electrical Engineering, University of Florida
Focus
Analytics · Applied AI · Computer Systems · Applied Probability
Experience
7 years professional poker · Applied probability
Venture
Co-founder, DMG Decisions
Building prediction and decision infrastructure for businesses.

Backstory

I find predicting the future to be incredibly satisfying. Predictions are good indicators of how much you know about the world, but more importantly of how much you know about yourself.

I sold sodas out of my backpack in middle school and washed dishes professionally in high school. I studied electrical engineering at the University of Florida — not because I wanted to be an engineer, but because I wanted to be literate enough to build what I saw in my head.

After college I played professional poker for seven years. In late 2019, while backed by ImAWhale, then a poker-staking group, I generated more than $25K for the stake over roughly three months. That's where I learned to separate edge, backing, and risk of ruin.

Get in touch with Andrew →

The Bets

Part of that poker bankroll later became an early Solana position in 2020 — turned a shoestring into a meaningful return, then watched most of it evaporate. That experience reinforced what the poker table had already taught me: finding an edge and keeping the money are two very different skills.

As CTO of DMG Decisions, I'm building prediction models with machine learning for businesses. Separately, I'm running an independent R&D project on computer vision for sports — the proving ground for physical-world sensing.

See the CPR prediction model →

Sports is the proof of concept. Physical-world sensing is the market.

Where This Goes

The same discipline extends to every domain where humans generate unpriced signal.

// selected_systems

LIVE + R&D, BUILT AND TESTED

CV Pipeline — Proprietary Data Manufacturing

Computer vision for extracting signals that don't exist in any dataset. CUDA-accelerated inference on local GPU hardware — extracting Table Proximity Index, Serve Toss Variance, and Split-Step Latency from free sports broadcasts. Proprietary data manufactured from public video for the cost of electricity.

The moat isn't the model. It's data that doesn't exist until you build the sensor.

Along Came Polly
Bayesian decision engine that sizes bets using Kelly Criterion — the same math Ed Thorp used to beat the market.
brier0.2243
kelly_f*0.34
tests136/136
PythonBayesianKelly
R&D

Along Came Polly — Decision Engine

A 4-phase decision optimization engine — 136 passing tests, pre-seed R&D stage. Kelly Criterion bet-sizing, Bayesian updating, calibrated probability outputs. The framework is domain-agnostic: any domain where you can measure an edge and size a position.

DMG Decisions ↗
Prediction and decision infrastructure for businesses. Projects include retail demand intelligence (179K transactions, Two-Stage Hurdle Model), autonomous sports prediction, and a Bayesian decision engine.
projects3
txns179K
modelslive
ConsultingML InfraProduction
LIVE

Recursive Agentic Systems

Cross-project orchestration infrastructure that improves itself over time. Autonomous research loops that discover, synthesize, and distribute insights across every active project — sports prediction, demand forecasting, computer vision — without manual coordination. Inspired by Karpathy's recursive research methodology. The system that builds the builder.

CPR Model ↗
Autonomous prediction daemon polling every 5 minutes. $20K of my own capital deployed, up $4,304 (+21%) in live trading. Dual-instance canary architecture with Brier-scaled Kelly sizing and 4-gate defensive filter. Purely statistical — CatBoost residual model on historical match data, no computer vision.
capital$20K
pnl+$4.3K
roi+21%
CatBoostSQLite24/7
LIVE
CV Pipeline
Most sports prediction models rely on structured data; this pipeline extracts signal directly from broadcast video. YOLO11m-Pose on local GPU, homography mapping, and Kalman-filtered kinematics have processed 465K frames across 9 batches, surfacing Table Proximity, Serve Toss Variance, and Split-Step Latency from free broadcasts.
modelYOLO11m
frames465K
batches9
VisionCUDAR&D
R&D
NanoClaw REPL ↗
Won the NanoClaw challenge in the Everything Claude Code repository (180K+ stars). Zero-dependency AI agent REPL with persistent sessions, dynamic skill loading, and markdown-as-database storage. PR #233 merged to main.
stars180K+
tests14/14
lines599
Node.jsOpen SourceAI Agents
LIVE
2015-2018
B.S. Electrical Engineering
Late 2019
ImAWhale $25K+ Result
2024-Present
DMG Decisions
Feb 2026
APOLLO Live
Apr 2026
CV Pipeline

The Infrastructure

Five-stage pipeline from raw sensor data to sized position. The same DAG processes sports telemetry today and retail biometrics tomorrow.

PREDICTION_PIPELINEARCHITECTURE SNAPSHOT
RESEARCH
hypothesisdata_auditedge_scan
INGESTION
sensor_telemetrypos_datamatch_api
MODEL
feature_engCatBoostcalibration
VALIDATION
CPCV(5)purge=12hembargo=6h
EXECUTION
kelly_f*=0.34half_kellyrisk_gated
TX_QUEUE
BRIER0.2243
KELLY_F*0.34
SHARPE_OOS1.21
TESTS136/136
LATENCY12ms

What I'm Exploring

Computer Vision & CUDA Extracting fatigue signals from sports video, pose tracking, optical flow, GPU-accelerated inference
Recursive Agentic Systems Self-improving orchestration loops, cross-project instinct promotion, Karpathy-style continuous learning
Niche Market Prediction Autonomous daemons hunting mispricings, Kelly-sized portfolio management, cross-domain arbitrage
Agentic AI Orchestration Multi-agent content generation, 100+ tool scaling, spaced learning pipelines, latency optimization
Decision Optimization Bridgewater-style institutional decision-making applied to time allocation, Bayesian updating, NPV analysis
"The first principle is that you must not fool yourself — and you are the easiest person to fool." — Richard Feynman

Let's build something at the frontier.

ACCEPTING CONNECTIONS