I build agentic systems, evaluate foundation models for Australian operating contexts, and make tools for creative work. The common thread is bounded systems, evidence, and keeping people in control.

Leaderboards are a starting point. I am interested in repeatable tests tied to local tasks, constraints, cost, and evidence.

• hexapla: early evaluation harness for running repeatable model tests, scoring outputs, and producing comparable result matrices.
• react-harness: instrumented ReAct loop for comparing LLM coding agents on multi-file tasks, measuring success, tool-call efficiency, error recovery, and dollar cost. Includes a small GLM-5.2 vs Claude Opus 4.8 run, with caveats stated.
• voxtral-bedrock-sydney: local web app demoing Mistral Voxtral on AWS Bedrock in ap-southeast-2, covering transcription, audio Q&A, summarisation, and translation with data kept in Australia.

Sample code I built in my role at AWS. These are reference implementations, not production systems, and the views here remain my own.

• sample-agentcore-delphi: learned cross-model router on Amazon Bedrock AgentCore Runtime. It picks a foundation model per request across families such as Nova, Llama, Gemma and Claude, and ships a test harness that checks whether routing beats a single model on quality and cost, and reports when it does not.
• sample-kiro-enterprise-governance: a Kiro Power for enterprise compliance workflows: standards enforcement through Bedrock Knowledge Bases, DOCX/PDF to markdown and back for spec-driven development, and a jurisdiction-agnostic AI assurance walkthrough.

I use AI and automation to help people inspect, transform, and finish work they already care about. The source material and creative decisions remain human. Bullpen covers the writing side; these projects cover music.

• asian-sentry-techniques: orchestration and composition notes from working in Logic Pro with foundation models. AI as co-composer, not autocomplete.
• music21-mcp: symbolic MIDI analysis and editing operations exposed through MCP, built on music21. It supports key and chord analysis, transposition, quantization, melody extraction, structural heuristics, and basic harmonization.
• preset-semantic: local-first semantic search for music-production presets across Helix, Arturia, and Decent Sampler. It turns musical descriptions into loadable preset candidates through a CLI or MCP tool.
• reascript-mcp: tools for reading REAPER project state and generating Lua ReaScripts for actions that can be checked inside REAPER.

Local-first utilities for narrow problems, and software for the clubs and people around me.

• opencourt-display: open-source Raspberry Pi signage for sports clubs: court allocations, honours boards, and notices on clubhouse TVs. Volunteers keep using Google Slides or uploaded images. Started at a Melbourne tennis club; club specifics stay configurable and private.
• dv-joiner: sorts timestamped DV clips and creates grouped, deinterlaced H.264/AAC access copies with FFmpeg. The original DV files remain the preservation masters.
• simple-redact-video-tool: local web app for blurring or blacking out regions of video clips, rendered with FFmpeg.
• rain-cast: Android rain prediction for Australia using Bureau of Meteorology data.
• simple-guitar-tuner: Android instrument tuner built with AI-assisted development.

Headless autonomous loops run on a Mac mini, with workers that produce, a supervisor that triages, and Signal for delivery. No dashboards or polling.

• supervisor: two-layer producer and verifier architecture. Workers write run files; a separate triage agent classifies them and signals only when something matters. It can integrate with any tool that writes to the filesystem, including Obsidian, CI, scrapers, and task managers.
• bullpen: profile-driven writing pipeline with bounded OpenCode stages, role-specific model routing, structural validation, and rollback on rejection. Optional Claude Code agents support a more flexible interactive workflow. The human editor controls publication.
• podcast-transcribe-exec-brief: polls podcast feeds, transcribes episodes locally with faster-whisper, and writes exec briefs into an Obsidian vault. There is no database; note frontmatter is the state store.
• podcastindex-mcp-server: Podcast Index search, episode lookup, and feed discovery exposed as MCP tools.

I write about AI deployment in Australian regulated sectors, sovereign infrastructure, and building with agents at craiglawton.com. I am AWS-affiliated; the writing is personal.

• Blog: craiglawton.com
• LinkedIn: linkedin.com/in/craiglawton
• Podcast: Building a Better Geek