M&E Engineer / Systems Architect
Fixing systems in real life and building systems online.
Infrastructure Design Principles
Signal extraction from high-volume, low-signal data streams — RSS feeds, chat logs, knowledge graphs. Threshold tuning and relevance scoring strip out the churn before anything expensive runs, so downstream systems only ever see what matters.
Per-purpose model routing, batch processing, and tiered validation keep expensive compute to a minimum. Each task is matched to the cheapest tool that can do the job well, and validation escalates only when the payoff justifies it.
A cascading LLM architecture tries the cheapest model first and escalates only when confidence is low. Desirepath algorithms let concepts form organically from real usage, so structure emerges where it is actually needed instead of being imposed up front.
Key Algorithms
A task enters at the cheapest capable model and only escalates when confidence is low — most work never touches an expensive model.
Concepts form organically the way footpaths do — from repeated use, not top-down design.
Flagship Projects
A local-first thinking partner built on a personal knowledge vault — per-purpose model routing, batch curation scoring, and 3-tier validation, Dockerized.
A weekly newsletter pipeline that scrapes RSS feeds, filters for relevance with cascading LLMs, and generates a formatted PDF — noise reduction via threshold tuning.