I'm Kit, an engineer in Vancouver. I build LLM pipelines and agents, the serving layer under them, and the low-latency infrastructure and data systems that trading runs on.
7 projects, each with its own write-up. The round link on the picture goes straight to the live site.
2026
An AI analyst I run myself: every hour it reads 127 sources, runs a nine-stage LLM pipeline with a fallback model behind every stage, and writes one page of market conclusions that trace back to their posts.
Showing project 1 of 7: LLM-Digest
What the projects above are written in, the AI tools I build with every day, and the infrastructure I run, from a Proxmox homelab to servers in several regions. Grab a ball and spin it.
LLM pipelines and agents I run, the serving layer under them, and the tools that let other agents use my data. The LLM-Digest case study replays one real run.
A nine-stage pipeline, a RAG agent with cited answers and one model layer with fallbacks, running every hour.
Case study4-bit KV and block-sparse Triton kernels under a learned vLLM scheduler on one RTX 3090.
Case studyOpen-sourced TypeScript and Python MCP servers that give AI agents RootData's crypto data as tools, over stdio, SSE and Cloudflare Workers.
GitHubQuantized 27B models and vision models on a single 24 GB card with llama.cpp, Ollama and vLLM.
Notes
I'm in Burnaby, BC, on Pacific time.
Photographs from Wikimedia Commons: fog in the Savoureuse valley by Thomas Bresson (CC BY 3.0); Mystic Beach, Vancouver Island by Michal Klajban (CC BY-SA 4.0).
