Building a zero-cloud push-to-talk dictation daemon for macOS that captures CoreAudio, runs sub-100ms MLX Whisper inference on Metal, and injects prompts with automated Return key delivery across any focused window.
01. Writing & Research
Bridging Discord and iMessage to Antigravity CLI via a 120-line bash socket loop, local SQLite polling in ~/Library/Messages/chat.db, and Apple Silicon MLX Whisper voice transcription.
An empirical N=20 benchmark showing that removing role-play personas and anthropomorphic guidance from ReAct agent system prompts increases unit-test accuracy by 20% and reduces token usage by 23%.
Why running shell commands to call web APIs causes slow execution, wasted context tokens, and escaping bugs, and how in-memory MCP servers with TLS connection pooling deliver sub-second tool execution.
Why coding benchmarks reflect the prompt scaffolding, edit protocol, and feedback harness rather than raw model capability in isolation.
Using Level-1 git sandboxes for sub-millisecond rollback on failed edits and Level-4 meta-supervisors to break stagnation loops in coding agents.
Simplifying web applications using modern browser standards, vanilla JavaScript, and Bun HTTPS servers without build steps or complex toolchains.
02. Open-Source Systems & Tooling
Modular developer tooling and agent infrastructure built for declarative execution, repeatable benchmarks, and low latency.
apicat
YAML → Tool CallingAPIs as executable definitions. Compiles human-readable YAML specifications directly into TypeScript libraries and CLI tools for token-efficient LLM tool calling.
When to use:
When an agent needs compact, token-efficient YAML schemas for HTTP and WebSocket tool calls instead of verbose SDK wrappers.
prolific
Autonomous Coding Agent
Minimal, environment-scoped coding agent runtime in Bun with apicat tool routing, scoped persistence, and /bench harness.
When to use:
When building or testing coding agents that require isolated memory (SYSTEM.md, MEMORY.md, tools.yaml) per workspace.
benchforge
Evaluation InfrastructureBenchmark infrastructure and repeatable evaluation workflows for AI systems with deterministic verification and pass@k metrics.
When to use:
When benchmarking LLM reasoning, coding, or tool use with deterministic verification (compiler exit codes, test passes).
llm-scorer
LLM Judge PrimitivesPrimitives for scoring model outputs, multi-metric grading, confidence scoring, and step-by-step chain-of-thought verification rubrics.
When to use:
When building automated evaluation pipelines requiring position-bias mitigation and rigorous judge rubrics.
compare-llms
Reasoning DivergenceTools for empirical model comparison, probing reasoning trajectories, and analyzing output branch divergence across model architectures.
When to use:
When visualizing reasoning branch paths and comparing outputs across model architectures.
vanilla-light
Zero-Build Web StackA no-build, dependency-free full-stack web framework on Bun with reactive client-side Proxy bindings and native DOM standards.
When to use:
When developing web apps without bundlers or build steps using Bun HTTPS + reactive DOM standards.
gateway-agent
Personal Messaging Gateway
Lightweight daemon bridging Discord and iMessage directly to Antigravity CLI via Unix pipes, SQLite polling in ~/Library/Messages/chat.db, and Apple Silicon MLX Whisper transcription.
When to use:
When you want to invoke and interact with local coding agents directly from mobile messaging apps (iMessage/Discord) with zero cloud dependencies.
03. Working Principles
- Ground all reasoning in verifiable execution. Language models generate plausible text; compilers, linters, and unit test suites provide ground truth. Agent reasoning loops must always be verified against deterministic runtime feedback.
- Use the smallest useful abstraction. Frameworks accumulate technical debt and obscure failure modes. Prefer declarative schemas, known APIs, Unix primitives, and inspectable intermediate states.
- Empirical benchmarking before optimization. Never tune prompts or agent architectures based on intuition alone. Build automated evaluation loops, measure baseline variance, and iterate methodically.
04. Agent Interfaces & Machine Endpoints
Standardized endpoints for autonomous AI agents, crawlers, and programmatic scrapers.
| Endpoint | Format | Purpose / Description |
|---|---|---|
| /llms.txt | Markdown |
Condensed index and systems directory optimized for AI LLMs & agents |
| /llms-full.txt | Markdown |
Comprehensive full-text context of all writing, architectures, and principles |
| /agent-instructions.md | Markdown |
Operational execution contracts, when-to-use matrix, and tool calling guidance |
| /index.md | Markdown |
Pure markdown representation of homepage (HTTP Content-Negotiation fallback) |
| /sitemap.xml | XML |
Canonical URL manifest and crawl directives for search engines and scrapers |