Engineering notes, project updates, and technical deep-dives. [ RSS feed ]
How ai-raccoon 1.53 moved its embedding model from the GPU to the Apple Neural Engine on an M4: 77 mJ per chunk instead of 283, half the time, faster searches. What it took to get there (a re-exported fp16 model in Apple's ANE layout, published on Hugging Face, and a background CoreML compile) and a plain-language primer on NPUs.
[ Read more → ]How ai-raccoon 1.44.3 to 1.51.2 cut idle CPU by about 90%, shrank the server from 6.8 GB to 1.75 GB and moved embedding to the GPU, with a plain explanation of ONNX, ONNX Runtime execution providers, WebGPU, MLX and why transformers belong on a GPU.
[ Read more → ]Pi is the TUI coding agent I run my multi-agent work through. Seven open-source extensions I built for it: delegation, monitors, fallback, cron, MCP, signals, keys.
[ Read more → ]Five architectural changes from v1.6.0 to v1.33.2: proxy server, event pump, custom embeddings, code corpus, project identity.
[ Read more → ]How AI agents built and power a full-stack job-search assistant — from CV tailoring to interview prep — and what that says about what's possible when agents write the code and AI runs the product.
[ Read more → ]How combining Semantica's causal knowledge graph with ai-badger hooks and ai-raccoon's persistent vector+BM25 memory layer automates the creation and retrieval of auditable decision chains for AI coding agents.
[ Read more → ]AiRaccoon is a local-first MCP memory server for AI agents: encrypted SQLite, hybrid FTS5+vec0 search, project and workspace isolation, cloud sync. Measured on 2026-08-04: 60% top-3 match rate on the 35-query baseline, 6/6 section hits on the structure-aware harness.
[ Read more → ]A tour of the skills my agent framework ships — what each one is for, what it actually does to your repo, and the situation that makes you reach for it. Starting with the one I open almost every day.
[ Read more → ]Part 2 measured a month of AI agent costs. This is what I changed because of it: my agent framework now picks the right model for each job, refuses hand-offs that name no model, and has a new team-lead role whose only job is routing work.
[ Read more → ]Part 1 compared three agents over two weeks. Part 2 reads 24 days of ccusage data off one machine — $10,453 in API-equivalent estimates, 97.81% cache reads, and the kinds of sessions that actually raise the bill.
[ Read more → ]A practical look at code-review-graph — how it builds a structural knowledge graph of your codebase, what it actually found in two real projects, and whether the token savings are real.
[ Read more → ]Data-driven comparison of Claude Code, Hermes Agent (via OpenRouter), and GitHub Copilot with real token usage, cost breakdowns, and the 80/20 strategy.
[ Read more → ]A practical review of Hermes Agent — how it compares to Claude Code, its hook system for prompt customization, token-friendly skill design, and why it pairs well with ai-badger.
[ Read more → ]What ai-badger is, how the stack×feature catalog model works, and why portable agent knowledge matters when you run Claude Code, Hermes, Copilot, and Junie side by side.
[ Read more → ]