~ / writing
The workshop log.
Not a blog. Architecture decisions, failure analyses, and research notes from systems that actually run. Written while the soldering iron is still warm.
agents & architecture
Aegis Falls: Frontier-to-Local Agentic Cascade
The complete system — frontier reasoning falling through orchestration into local execution, held together by cognitive memory. How all the layers connect and why each exists.
Design Review With a Panel of AIs
Conference rooms with locked contracts where agent panels argue designs into shape — catching in minute four what would cost a weekend at hour twenty.
Building Self-Improving Agents
Six iterations of an agent platform and what each one taught: agents don't improve by making the model smarter — they improve by building better infrastructure around it.
cognitive memory
ACT-R Memory: March vs Now
Five months of production: the black-hole bug health checks missed, running ACT-R and mem0 in parallel, and a custom context-injection plugin with an audit trail.
The Memory Layer: When an AI Actually Remembers
ACT-R in production: activation dynamics, spreading activation, logarithmic decay — and what changes when your AI stops forgetting you between sessions.
ACT-R Cognitive Architecture for AI Agents
A human-cognition architecture, adapted for autonomous agents: declarative chunks competing on activation, procedural rules learning utility, working memory buffers.
product & scoring
The Catch Log Is the Reviewer
Two times the forecast model got caught lying — flattering averages and textbook curves — and the multiplicative-modulator rebuild that fixed both, pinned by regressions named after the embarrassments.
autonomy & verification
Agents That Build While You Sleep
Unattended builds that survive past hour three: checkbox plans with proof commands, fresh context per step, fail-twice-revert discipline, and ten laws instead of a rotting manual.
My Optimizer Kept Picking Lucky Fools
A tournament with noise in it is a lottery, and optimizers always enter. How demanding rematches separated genuine champions from strategies that had one good day.
My Tests Never Met My Code
Tests written after and independently, blindness enforced with chmod, searcher and judge split between agents — because a model grading its own homework passes everything.
practical guides
Self-Hosted Guides
The how half of the workshop log: service skeletons, systemd discipline, honest watchdogs, snapshot policy, multi-provider routing. Copy-paste-able, reboot-proof.
cost engineering
Secondary Agent Economics
~8 billion tokens of agent work in a month for under $60: a mid-tier primary model, an idle subscription's second life, and free previews as burst capacity.
infrastructure & opinion
Why Local AI Matters
A manifesto: most people rent their intelligence. On owning your inference stack, the structural risks of API dependence, and the rule that started it all.
Homelab Architecture for AI Development
Why Proxmox over bare metal, how GPU passthrough earns its keep, storage and network decisions — every choice traced back to a specific problem that demanded it.
$ tail -f ~/writing/*.md — new entries when something ships or breaks.
suggest a topic →