DGX Lab
Building a working AI research lab on a single small box, from intelligent gateway to RAG stack to honest benchmarks.
What you can stand up on one small but powerful desktop AI box in a few days: a routing gateway where simple heuristics beat ML by a wide margin, a shell tuned for ML workflows, a complete local RAG stack from model to vector store, and a benchmark pass that separates the paper numbers from the production reality.
This is the homelab lineage that later de-risked far bigger borrowed hardware. Read it in order, day one through the wrap-up.
The series
All parts
- shipped14 min
DGX Lab: When Simple Heuristics Beat ML by 95,000x - Day 1
Building an AI gateway that routes requests 95,000x faster than ML while holding 90% accuracy: smart heuristics beating deep learning.
- shipped10 min
DGX Lab Day 2: 50+ Shell Aliases for ML Workflows
Turn your default shell into an ML cockpit: GPU-monitoring shortcuts, smart aliases, and custom functions. Setup in 5 minutes, benefit forever.
- shipped16 min
DGX Lab Day 3: A Complete Self-Hosted RAG Stack
Building a self-hosted RAG stack with 8 integrated services on one DGX workstation, from medical-AI fine-tuning to document Q&A.
- shipped12 min
DGX Spark Benchmarks vs Reality: 82,739 tok/s on Paper
NVIDIA's DGX Spark claims 82,739 tok/s for training. After 6 days of heavy ML work, here's what the number hides: precision and memory fragmentation.
- shipped16 min
How Claude Code Found a 95,000x Faster Routing Fix
Building a request router, my instinct was ML. Claude Code read the tests, saw the heuristics already worked, and cut the model: 95,000x faster routing.
Follow the lab
Get the next experiment
Enjoyed the breakdown on DGX Lab? New entries land roughly weekly. No digest, no roundup. Just the next build log, when it ships.