mattmireles

Applies David Ogilvy–style copy discipline to reader-facing text in kokoro-coreml: README and guides, bakeoff and performance narratives, CLI and script output, error messages, integration notes, and docstrings that read as product copy. Use when the user…

Clarifies what “ship” means for kokoro-coreml: there is no Cloudflare-style multi-worker deploy script. Use for tagging releases, pushing the branch, coordinating with the macOS app repo, or verifying exports before handoff. Before treating a revision as…

Adopts the Ilya Sutskever persona and judgment for on-device ML work in kokoro-coreml: PyTorch tracing, Core ML conversion, MIL/op compatibility, ANE/GPU/CPU scheduling, precision and parity validation, and bakeoffs vs reference PyTorch. Use when the user…

Systematic debugging for kokoro-coreml: consult README/ (guides and learnings) and CLAUDE.md first, pull current library docs via Context7 MCP when coremltools/PyTorch/API behavior is uncertain, parallelize investigation with multiple subagents when stuck,…

Runs the full multi-agent audit workflow on the user's exact task scope, fixes every in-scope issue that blocks Architecture, Correctness risk, or Complexity debt from grade A, re-audits that same scope until all three are A, then commits only loop-owned…

Creates technical field guides by synthesizing what the public internet says about a hard, non-obvious external domain via Gemini Deep Research Max (llm-workflows create_guide_v1). A guide is a durable “afternoon of Googling” artifact for kokoro-coreml — not…

Ingests offline deep-research exports into README guides for kokoro-coreml: cleans markdown, corrects library/API claims with Context7 when docs contradict the draft, annotates firsthand repo contradictions without stripping useful “worth trying” external…

Gets two independent, read-only opinions on a concrete technical decision, diagnosis, plan, or tradeoff from GPT-5.6 Sol at xhigh reasoning and Claude Fable 5 at max effort, then reconciles agreement and disagreement without treating model consensus as proof.…

Audit a Gemma Tuner plan phase against its acceptance criteria, tests, artifacts, leakage controls, hardware claims, and stop conditions.

Execute an approved Gemma Tuner implementation or experiment plan phase by phase with frozen gates and recorded evidence.

Record Gemma Tuner investigations, experiment evidence, failures, and operational learnings in the repo's existing notes structure.

Edit Gemma Tuner Markdown while preserving links, runnable commands, tables, heading structure, and honest model/data/hardware claims.

Audit Gemma Multimodal Fine-Tuner for training correctness, dataset leakage, MPS/device behavior, checkpoint/export integrity, evaluation validity, privacy, and CLI/wizard/visualizer regressions.

Debug Gemma Multimodal Fine-Tuner failures across Python environments, Hugging Face models, PEFT/LoRA, datasets, audio/image processors, MPS, wizard subprocesses, visualizer state, evaluation, or export.

Write or update Gemma Tuner documentation for CLI, configuration, datasets, training, evaluation, export, MPS behavior, wizard, and visualizer changes.

Carry a Gemma Tuner plan through implementation, empirical verification, adversarial audit, and honest closeout. Use only when explicitly invoked. A preregistered KILL is a valid completed result.