A working catalog
of AI tools,
built eval‑first.
AI tools I build to one rule: the eval comes before the feature. None of them make the call for you. Every one keeps a human as the judge. Amplification, not replacement. You stay smarter than your tools.
I start with the eval.
Plant the failures, strip the answer key, gate the build in CI, and keep a human as the judge. Two shipped repos, the same move on different problems: discovery, and retrieval/grounding.
Anton Scout
Can it tell a real idea from hype? A decoy-injection eval says yes: 1.0 decoy-catch (stable across runs) and 0.73 recall after self-consistency, on 27 candidates. CI-gated, and replayable without a key.
- Technique
- idea extraction + self-consistency scoring + decoy-injection eval
- Medium
- Open source · GitHub
Coherence Keeper
Does the answer really come from the source? A planted-contradiction eval takes F1 from 0.50 to 1.00 at a 0% false-positive rate. Stable across runs, and replayable without a key.
- Technique
- RAG (Cohere Embed + Rerank) + LLM judge + planted-contradiction eval
- Medium
- Open source · GitHub
In the world.
Shipped, running, or honestly in progress. The open-source entries link straight to source; nothing here is a mock-up.
- 02
Jellybook
A real comic reader for Jellyfin: CBZ/CBR, two-page spreads with wide-page auto-detect, right-to-left manga mode, progress sync, neighbour preloading. Replaces the single-page default reader without touching ebook support.
Shipping- Technique
- Jellyfin plugin · .NET 9 · web injection
- Medium
- Open source · GitHub releases
View source → - 03
Son of Anton
A three-machine Tailscale mesh (two GPUs plus an always-on hub) running a tiered local-LLM router. Each prompt goes to the smallest model that can handle it, and only escalates to a bigger local model, or the cloud, when it has to. It's what runs the inference and the overnight jobs behind the other projects.
Running- Technique
- tiered local-LLM routing + Tailscale mesh
- Medium
- Personal infrastructure · runs nightly
- 04
Reprise
Adaptive learning that works out what you actually know, topic by topic, from your real activity, then tunes what it shows you next. Tell it 'too basic' or 'too advanced' and it adjusts. I use the core loop every day; the standalone desktop app is still in progress.
In progress- Technique
- per-concept skill modelling + generation conditioning
- Medium
- Native desktop (Tauri + Rust) · local-first