VS Labs → Computer science and AI01 / 06
Computer science and AI internals, animated one step at a time.
A ferrofluid display: black liquid behind a backlit glass window rises into spikes as its field pulses, and higher as the page scrolls.
Lab → Scenes02 / 06
The first released scene arrives the evening its video posts. Until then the frame below shows what a scene is made of.
Lab → One step at a time03 / 06
What a scene is made of, before the first one lands. Each caption below is one step, and the frame is drawn live by the renderer that makes the videos.
Eight glass pucks on the mat. Nothing is lit yet; the scene rests.
One puck lights when a step is about it. The tube underneath marks the row.
Blue glass means settled: ruled out, finished with, still legible.
Pointers name positions. Every footnote states what was counted.
Lab → Record04 / 06
The engineer behind VS Labs builds full-stack systems for a living: services, data pipelines and cloud infrastructure that run all day. The lab is where the same habits meet AI: build it end to end, measure it, publish the numbers.
Three agents review a pull request at once (quality, security, documentation), calling Roslyn, Semgrep and GitHub through a custom MCP server. An orchestrator merges their findings into one ranked review in which every claim names its source.
C# · .NET 10 · MCP · OPENTELEMETRY
Upload documents, ask in plain English, get answers that cite their sources. One configuration value moves the entire stack between local models with a local vector store and a cloud pair, and a comparison view runs both side by side.
C# · .NET 10 · REACT · OLLAMA · CHROMADB · AZURE AI SEARCH
It predicts equipment failure from sensor data, then watches its own inputs. When traffic from an unseen operating regime drifted, it retrained, evaluated and registered a challenger, and a human approved the promotion. No cloud secret is stored anywhere.
PYTHON · XGBOOST · MLFLOW · EVIDENTLY · AZURE · BICEP
Can a fine-tuned 7B to 14B open model replace the frontier model inside the code reviewer at the same precision and recall? LoRA training and serving on the lab’s own hardware, with a test set touched exactly once.
LORA · VLLM · MLFLOW · DVC
In progress → the results table is empty by design
Lab → How it’s made05 / 06
Every scene is a file in a public repository. The lab validates it, renders it and encodes it locally, and this section’s frame is that pipeline, drawn by the same renderer.
Two NVIDIA DGX Sparks with 128 GB of unified memory each, bridged over ConnectX-7. Models are served and fine-tuned here, and the numbers below were measured here.
$ vllm bench serve --random-input-len 1000 --random-output-len 256
tensor_parallel_size ....... 2
resident model ............. 320B MoE · 4-bit
memory in use .............. 244 / 261 GB
first token, 1k prompt ..... 1.1 s
decode floor ............... 13.5–15.2 tok/s// DECODE FLOOR ON RANDOM TOKENS · REAL-TEXT RUNS PENDING · 2 × 128 GB UNIFIED
// ONE SPARK · 10-QUESTION DOCUMENT Q&A BENCHMARK · MEAN END-TO-END ANSWER · 2026-07
Lab → Next06 / 06
Commission a scene for a system that needs explaining, or follow the lab as it traces the next one.