$ cat ./technical-notes

Technical notes.

Field notes on local LLM infrastructure, coding-agent workflows, evaluation loops, and the places where language models should hand authority back to deterministic code.

6 notes

local AI · operator notes · field note

Local LLM Operator Workbench

A practical local-first coding-agent stack built around LM Studio, model presets, and MCP.

Operational notes for running local coding agents: LM Studio as an OpenAI-compatible endpoint, task-specific presets, MCP scope, and context budgets that keep edit loops responsive.

  • LM Studio
  • Local-first
  • MCP
  • Qwen
  • Gemma
note ↗
agent workflow · routing · systems note

Coding Agent Model Routing

Plan/act separation, local fallbacks, and choosing the smallest capable model for each loop.

A compact routing doctrine for coding agents: planning versus acting, local versus hosted, fast versus deep, and when to split a task instead of feeding one giant context window.

  • Plan/Act
  • Model routing
  • Context budgets
  • Local-first
  • Workflow
note ↗
coding agents · operations · case study

Cline Ops Discipline

Workflow rules for keeping coding agents narrow, reviewable, and out of runaway context.

The operating model behind using Cline for repo work: repo questions, plan-edit-verify passes, broad-task discovery, and monitoring for tasks that start to sprawl.

  • Cline
  • GLM
  • Repo workflows
  • Large tasks
  • Verification
note ↗
evals · coding agents · lab note

Cline Evaluation Lab

A local exploration of smoke tests, contract tests, and agent regression signals.

A look at an evaluation stack for coding agents: provider smoke tests, transform contract tests, end-to-end tasks, metrics, and reporting patterns for catching regressions.

  • Evals
  • Smoke tests
  • pass@k
  • Regression testing
  • Cline
note ↗
operator tooling · telemetry · prototype note

AI Observer

A lightweight terminal telemetry experiment for command, error, and repo-state context.

A small local observer pattern: shell hooks capture command/error metadata and snapshot repo state so an assistant can understand what changed without relying on shell history.

  • Shell hooks
  • Telemetry
  • Git snapshots
  • Operator context
  • Local only
note ↗
AI systems · safety boundary · architecture note

LLM Trader Runtime Boundaries

LLMs propose trades; deterministic services validate, gate, and sign.

A note on the runtime boundary in LLM Trader: committee agents can generate proposals, but the risk gateway owns policy, quote freshness, audit records, and signing.

  • LLM Trader
  • Risk gateway
  • Deterministic policy
  • Shadow lanes
  • Audit
note ↗