Blog

Consilience is a knowledge game where you assemble real, cited facts into cross-disciplinary insights and defend them to an AI tutor — but the pass/fail verdict is computed from the citation graph, not the model. A fluent answer with no sources cannot win. Design notes on separating the layer that talks from the layer that judges.

I run one local chokepoint for every AI call I make, with data sorted into tiers and each tier cryptographically limited to the providers I trust with it. Sensitive data reaching the wrong model is enforced against, not hoped against. Here's the architecture and what proving it taught me.

A structured, fully-cited research project covering all 16 regions of New Zealand — not prose, but a typed entity graph of problems, claims, drivers and rival camps, where every quantitative claim has to cite a source or the build refuses to render.

A small Bitwarden helper that finds duplicate logins and reports them — and deliberately never deletes a thing. The interesting decision wasn't the dedup logic; it was refusing to touch the vault.

A one-page panel that auto-discovered my local projects and launched their servers on click. It worked — and then I built something that did the job better, so I retired it. A short note on a tool that earned its replacement.

A live EQ + compressor + limiter for Windows audio, built like one channel of a mixing desk — so whispered dialogue and explosions land at the same volume. Python, a virtual cable, and a few honest trade-offs.

A Tauri desktop app that starts, stops, health-checks, and embeds every local service I run — and the place this site's status-first design language came from. Includes the Windows process-management rabbit hole.

An ad-free, New Zealand-resident alternative to the big-platform stack — assembled from unmodified open-source apps on my own hardware. The vision, the assemble-don't-fork rule, and why the licensing actually works.

Five prototypes trying to turn a knowledge base into a video game — Consilience, Tickscape, and three single-file concept toys — and the design finding that survived all of them: the concept has to be the toy, not the test.

How a 7B model on an 8 GB consumer GPU got near-14B answer quality, fully offline, with zero observed hallucination on spot checks — by splitting knowledge authoring from knowledge synthesis. There's a live demo.

A layer-by-layer walkthrough of the Job Scout scoring loop — from Python function call to transformer forward pass — plus what the replacement dashboard does instead, and why the second approach is more honest about what each tool is good for.

A post-mortem of scout_mvp.py — what I built, why I thought local-AI judgment would work, what actually broke, and what replaced it.

A zero-shot voice clone built from one short reference clip of my own audio — how it works, how convincing it actually is, and where I'll use it.

A home-automation design built around what I want my house to do — and, just as importantly, what I don't want it to do. The principles are locked; the hardware isn't on the network yet.

I built a personal finance dashboard around the Akahu open banking API to stop manually reconciling bank exports. Node and React — which I'd choose differently today. The Akahu integration is the leverage that actually matters.

A Windows file-organisation system built around the rule that no automation gets to touch my filesystem without a dry-run first. Everything + Python organize + digiKam, with JSONL logs of every move.

A local Flask app combining NZ politician profiles with a combined Vote Compass / Political Compass questionnaire mapped against the actual platforms of NZ parties. What I built, what I left out, and why.

An NZ-specific streaming aggregator that combines TMDB provider data with sitemaps from TVNZ+ and ThreeNow — so 'free to watch right now in New Zealand' actually means something. FTS5 search, sitemap parsing, and what's still ahead.

A Bitcask-style key-value store with a Rust core and PyO3 Python bindings — designed to teach me Rust the hard way. I retired it. Here's why, and what I'm doing instead.

This week's digest covers two threads from the arXiv cs.AI feed: advances in self-reflective multimodal generation and continual learning, and the case for engineering discipline in personal AI agents.

Three threads from this week's reading: safety scaling laws in clinical LLMs, structured control flow for long-horizon agents, and the disclosure norms AI is reshaping in security research.