Cybersecurity student. I care about protection: of people, of agents, and of the shared spaces they work in. Most of what I build runs unattended, so it has to be honest about what it did and why.
The lab — a small home lab built from ordinary machines, run like infrastructure: I am no expert. I hope what I build is of use to anyone, and if you have any suggestions please feel free to contact me.
- A workstation tier that does the heavy work — local LLM inference and a small fleet of cloned VMs for parallel data collection. It wakes on demand under reference-counted leases, sleeps when nobody needs it, and yields instantly when a person sits down at it.
- An always-on sentinel laptop: the lab's wake sender, host of a second-eyes reviewer agent, and a small local inference tier. It is the machine that is awake so the others don't have to be.
- Tiered local inference on both machines — bigger models where the hardware allows, small ones where it doesn't — with mechanical fallbacks when a tier is down.
- Storage discipline: free-space floors on every drive, verified backups, and bulk movement that is capped and scheduled so the lab never takes the household network down with it.
The harness — the lab exists to run an agent system, and the design rule that emerged is that commitment belongs to code, not to the model. A cloud agent head holds the judgment. A dispatcher layer (in shadow testing) routes bounded work to tools and local models: the model does language only — intake, summarization, phrasing — while routing, required checks, refusals, and verdicts are derived by code from verified findings. Every dispatch carries provenance and a gateway-issued idempotency key; gates fail closed. A separate reviewer agent grades the head's claims against the records, and an always-on local agent trained on my own corpus is in training as a lower-cost first responder.
The workloads
- Nightwatch — an agentic framework for building calibrated, falsifiable agent systems. Rules are pre-registered as code, so the system argues through falsification, not discretion. The reference build is a prediction-market research engine: quarter-Kelly sizing, mechanical EV gates, Brier-score calibration, kill conditions on the doctrines themselves.
- An audio-driven music-video engine: real audio features drive every frame. No procedural stand-ins.
- A wake chain for the lab itself: lease-based wake, an activity-aware sleep decider, and a power log that records every verdict.
- A congressional disclosure pipeline: STOCK Act filings ingested from primary sources.
- A collectibles pipeline: inventory classification and pricing for an established resale operation, Python + SQLite + local LLMs. (A companion resale venture is currently parked.)
Stack: Python, C++, SQLite, Next.js, local inference (Ollama), VM fleets, multi-agent orchestration.