quality systems · on-prem AI · measured

Coinupbtc

Quality systems on a plant you can walk. Models that never leave the box. Measured, not claimed.

Helix is the 90-second hiring walk. The rest is measured inference, open repos, and work made on this machine.

Two NVIDIA DGX Spark units, champagne metal bricks, linked as a cluster
hosted here — two DGX Sparks
Pseudonymous · hiring-friendly coinupbtc.com

The machine

Hardware behind the proof

GPU memory
121 GB
unified CPU+RAM, on-box
Storage
3.7 TB
NVMe, local weights
Decode
169 t/s
measured coding throughput
Art
on-metal
ComfyUI, no cloud API

Selected public work

What a hiring manager should open

all repos
Helix QMS Desk Monday war-room — sticker versus certificate mismatch, CAPA, training gaps

helix-qms-desk

Interactive ISO 13485 quality desk for a fictional biomedical plant. Monday: sticker ≠ certificate, CAPA 8D, supplier SCARs, fail-closed integrity checks. Synthetic records. Open it in the browser.

Endpoint-agnostic LLM benchmark suite — score dials and pass/fail checks

zwell-bench

Release gate for local LLMs — coding, vision, tool and agentic checks. A candidate ships only at 19/19. Caught a specialized build that beat the headline metric and still failed.

Local GPU fleet dashboard — graphs, model inventory, endpoint health

spark-console

Local GPU and fleet dashboard — graphs, model inventory, endpoint health, read-only service board. 121 GB unified memory. Nothing phones home.

Local LLM serving bakeoff — measured tokens and tuning curve

miaai35-tune

Flag-by-flag llama.cpp bakeoff on DGX Spark — serving settings from evidence, not vendor slides. Coding decode measured at 169 t/s.

Two NVIDIA DGX Spark units used for dual-node inference

dream-stack

Two-node NVIDIA GB10 inference: Dockerized vLLM tensor-parallel plus a llama.cpp roommate. Clone, copy .env, bring the stack up.

the lab

Browser mechanisms — memory packer, speculative decoding, render-cost timer. No account. No checkout.

Elsewhere

Find me in the open