Electrical Engineer | CEO & Founder, YOSO-YAi LLC
Electrician turned EE. I do surgery on open-weights models and measure Energy Per Intelligence on the hardware that runs them.
Fluor Corporation · Tesla Gigafactory · Meta Data Centers · YOSO-YAi
"The future of labor is compute."
I started as an electrician. I pulled wire on Fluor Corporation job sites across the country, learned instrumentation and controls at Tesla's Gigafactory, and wired Meta's NightCrawler data center. I have always thought in watts.
Now I apply that lens to AI models. Every token a model generates is a physical event — a specific quantity of energy flowing through a specific piece of silicon. The AI industry treats tokens as abstract units of text. I measure what they actually cost: joules.
I design every schematic, every PCB layout, and every enclosure for every product YOSO-YAi ships. Self-taught in KiCad 9.0 and Fusion 360. Every board starts on an exposed breadboard where every wire is visible, then moves to production PCB.
YOSO-YAi builds autonomous AI agents that replace human roles — not assist them, not augment them, replace them. Each agent is sold exclusively on custom Raspberry Pi hardware with an ATECC608B cryptographic wallet soldered to the board. No agent leaves without silicon.
Energy Per Intelligence (EPI) is a metric I created to fill a gap nobody else occupies:
EPI = Joules per Token / Task Accuracy
The people who do surgery on models never measure the energy cost of the result. The people who measure energy never touch the models. I do both — on the actual production hardware where the models run.
| Research Community | What They Do | What They Don't Do |
|---|---|---|
| Model Surgery (MoE-Pruner, NAEE, SparseGPT, Wanda) | Prune experts, remove heads, quantize weights | Measure energy. No joules/token. No edge hardware. |
| Energy Benchmarking (ML.ENERGY, EuroMLSys 2025) | Measure joules/token on stock models | Perform surgery. No model modification. No ARM. |
| YOSO-YAi EPI Pipeline | Surgery + Energy + Efficiency Ratio + ARM Hardware + Custom Instrument | This is the gap we fill. |
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ SURGEON │────▶│ PATIENT │────▶│ INSTRUMENT │────▶│ CALCULATE │
│ DGX Spark │ │ Pi Cluster │ │ epi-meter │ │ epi-bench │
│ GB10 128GB │ │ 4x Pi 5 │ │ RP2350 + CT │ │ EPI Score │
└─────────────┘ └─────────────┘ └─────────────┘ └──────┬──────┘
▲ │
│ FEEDBACK LOOP │
└───────────────────────────────────────────────────────────┘
DGX learns to predict EPI before deploying
The lab that powers everything. Four machines. Four AI employees. Zero overlap.
╔══════════════════════════════════════════════════════════════════════════════════╗
║ YOSO-YAi FACTORY ║
║ The Engine Behind the Product ║
╠══════════════════════════════════════════════════════════════════════════════════╣
║ ║
║ ┌─────────────────────────────┐ ┌──────────────────────────────────────┐ ║
║ │ DGX SPARK │ │ Pi 5 CLUSTER │ ║
║ │ GB10 · 128GB · 1 PFLOP │ │ 4x Pi 5 16GB · 64GB Total │ ║
║ │ │ │ │ ║
║ │ Autonomous Systems Eng. │ │ Data Engineer │ ║
║ │ ├─ NemoClaw (AI Employee) │ │ ├─ Academic DB Scraping │ ║
║ │ ├─ Model Surgery (EPI) │───▶│ ├─ Component/Company DB │ ║
║ │ ├─ Fine-tuning (NeMo) │ │ ├─ distributed-llama Inference │ ║
║ │ ├─ n8n Orchestration │ │ └─ EPI Measurement Target │ ║
║ │ ├─ Kalshi Trading │ │ │ ║
║ │ └─ Content Gen (ComfyUI) │ │ Orchestrator Pi (10" Display) │ ║
║ └─────────────────────────────┘ │ └─ Live Telemetry Dashboard │ ║
║ └──────────────────────────────────────┘ ║
║ ┌─────────────────────────────┐ ┌──────────────────────────────────────┐ ║
║ │ LENOVO M70q │ │ iMAC │ ║
║ │ i7-14700T · 32GB │ │ Alexandra's Station │ ║
║ │ │ │ │ ║
║ │ Hardware Engineer (Abner) │ │ 3D Printing Engineer │ ║
║ │ ├─ KiCad Board Review │ │ ├─ Enclosure Design Review │ ║
║ │ ├─ Schematic QC │ │ ├─ Bambu Lab H2D Control │ ║
║ │ ├─ BD Agent (Upwork) │ │ ├─ Prototype Manufacturing │ ║
║ │ └─ Senior Engineer Agent │ │ └─ Commodities Trading │ ║
║ └─────────────────────────────┘ └──────────────────────────────────────┘ ║
║ ║
║ ┌──────────────────────────────────────────────────────────────────────────┐ ║
║ │ EXPOSED BREADBOARD (Center of Lab) │ ║
║ │ Pico 2 (RP2350) · NTC Thermistors · PID Fan Control │ ║
║ │ Agent Heartbeat GPIO · UART Telemetry · Every Wire Visible │ ║
║ │ │ ║
║ │ Validates: WRK-IN BMC subsystems + epi-meter power measurement │ ║
║ │ Same silicon as production safety planes on WRK-IN & Controller │ ║
║ └──────────────────────────────────────────────────────────────────────────┘ ║
║ ║
║ Command: CEO >>> Autonomous Sys. Eng. (DGX) >>> Data Eng. (Pi Cluster) ║
║ Bambu Lab H2D ─── Prints all enclosure prototypes from CEO's designs ║
╚══════════════════════════════════════════════════════════════════════════════════╝
Each paper is a GitHub repository. README.md is the paper.
/data/has raw results./code/has reproduction scripts.
| # | Repository | Status | Core Contribution |
|---|---|---|---|
| 1 | energy-per-intelligence |
Upcoming |
Defines the EPI metric. Baseline measurements on Pi cluster. Introduces the epi-meter board. |
| 2 | epi-meter |
Upcoming |
Open-source power measurement board. RP2350 + 4x metering ICs + CT clamps. Full KiCad + firmware. |
| 3 | epi-bench |
Upcoming |
Shared tooling: EPI calculator, Pareto plotter, power trace logger, benchmark runner. Community CSV format. |
| 4 | expert-pruning-epi |
Upcoming |
MoE expert removal measured by EPI. Where is the efficient operating point? |
| 5 | mixed-quant-epi |
Upcoming |
Per-layer quantization evaluated by EPI. Two configs, identical perplexity, different joule costs. |
| 6 | attention-head-surgery-epi |
Upcoming |
Removing attention heads: does the energy actually drop, or do remaining heads compensate? |
Hardware-bound autonomous AI agents. No agent is sold without silicon.
┌───────────────────────────────┐
│ YOSO-YAi FACTORY │
│ DGX Spark · Pi Cluster · n8n │
│ Fine-tune · Surgery · EPI │
└───────────────┬───────────────┘
│
Models ship on hardware ▼
┌─────────────────────────┼─────────────────────────┐
│ │ │
┌──────▼──────┐ ┌──────▼──────┐ ┌───────▼──────┐
│ Community │ │ AGNT Board │ │ WRK-STATION │
│ FREE │ │ $1,800 │ │ $8,999 │
│ │ │ │ │ │
│ Fine-tuned │ │ 1 Agent │ │ 6 Agents │
│ models on │ │ CM5 16GB │ │ WRK-IN mgmt │
│ your own Pi │ │ 1TB NVMe │ │ Controller │
│ │ │ USB-C + Bat │ │ Silicon Moat │
│ YOSO-YAi │ │ Phone App │ │ 10" Touch │
│ Terminal │ │ ATECC608B │ │ Mesh + eSIM │
└──────────────┘ └─────────────┘ └──────────────┘
Four Custom Boards — All Designed in KiCad 9.0 by the CEO
| Board | PARD | Silicon | Role | Fabrication |
|---|---|---|---|---|
| WRK-IN | PARD-WRKIN-001 | CM5 + RP2350A BMC | System management: 6x TPS54202 power, PID thermal, 90-min battery, mesh, graceful shutdown. 44 connectors. | JLCPCB |
| WRK-AGNTs | PARD-AGNT-001 | CM5 16GB | Worker blade: 1TB NVMe, ATECC608B wallet (factory-locked), TMP117 thermal. 85x77.5mm, 6-layer. 6 per WRK-STATION. | PCBWay |
| WRK-STATION Controller | PARD-WRKC-001 | CM5 + RP2350 BMC | Owner's terminal: 10" touchscreen, dual-slot ATECC608B (cold vault + operational), eSIM + SX1262 mesh, vault mode. | PCBWay |
| AGNT Board | PARD-AGNT-002 | CM5 16GB | Standalone consumer product: USB-C + battery, display, phone app, single ATECC608B wallet. | PCBWay |
The Silicon Moat — 87 Cents. Five Locks. Physics, Not Code.
Every WRK-AGNT blade carries a Microchip ATECC608B — a $0.87 cryptographic chip that physically locks every agent's economic activity to the YOSO-YAi factory. The key was born inside the chip during manufacturing and can never come out.
| Lock | What It Does | If Compromised |
|---|---|---|
| Agent Identity | Unique key pair signs every transaction. Born in silicon. | Factory rejects signature. Transaction denied. |
| Encrypted Channel | Chip connects to one server only: YOSO-YAi factory. | No channel forms. Data cannot flow. |
| Factory Certificate | Factory fingerprint sealed into chip at manufacturing. | Factory rejects destination. Transaction denied. |
| Model Integrity | Hash of AI model. Detects tampering. | Wallet locks. Agent economically dead. |
| Factory Public Key | Factory's crypto identity burned into chip. | No peer to connect to. Communication impossible. |
Bypass requires replacing the chip — destroying the agent's identity, wallet balance, and transaction history. The cost of circumvention exceeds the cost of compliance. That is a moat.
The epi-meter Board — Open Source Power Measurement
A custom PCB designed to instrument the Pi cluster with dedicated energy metering hardware. The first YOSO-YAi board design that is fully public.
| Parameter | Specification |
|---|---|
| MCU | RP2350 (Pico 2 silicon) — same as production BMC |
| Energy Metering | 4x dedicated ICs (ATM90E26/ADE7753 class) over SPI |
| Current Sensing | 4x CT clamps (SCT-013 class) — non-invasive |
| Measurement | True RMS voltage + current, real power, power factor |
| Output | UART to Orchestrator Pi, 115200 baud, JSON |
| Measurement Point | AC side — captures total system draw including PSU losses |
Why not software measurement? The Pi 5 has no GPU telemetry. Software power estimation on ARM Linux is unreliable. An electrician does not trust a device's self-reported power draw. An electrician puts a meter on the circuit.
What gets published: Full KiCad schematic, PCB layout, Gerber files, complete BOM, RP2350 firmware in C, 3D-printed enclosure, assembly guide, calibration procedure. Everything needed to build one.
| Tool | Role in EPI Pipeline | License |
|---|---|---|
mergekit |
Model merging, DARE/TIES, expert manipulation | Open Source |
llama.cpp |
GGUF quantization, mixed-quant, Pi inference | MIT |
transformers |
Load/modify/save weights, head pruning, layers | Apache 2.0 |
safetensors |
Tensor-level weight surgery | Apache 2.0 |
distributed-llama |
Pi cluster inference engine | MIT |
NeMo |
Enterprise fine-tuning on DGX at FP8 | Apache 2.0 |
PEFT |
LoRA fine-tuning for energy payback analysis | Apache 2.0 |
2014-2017 Fluor Corporation Industrial electrical · Conduit bending · Journeyman
2017-2019 Independent Bitcoin trading · Project management (online coursework)
2020-2021 Tesla Gigafactory Instrumentation & Controls · Sensor integration · PLC logic
2022-2024 Meta NightCrawler Electrical Foreman · Power distribution · Data center wiring
2025-now YOSO-YAi LLC CEO & Founder · PCB design · AI hardware · EPI research
The future of labor is compute. Measured in energy per intelligence.
