SignalSift: my content intelligence daemon

modules/home/services/signalsift.nix. SignalSift follows the blogs, newsletters, channels and podcasts I care about, transcribes what has no captions, scores every segment against my interests with a local LLM, and serves the result on http://127.0.0.1:7878/ (a web UI and an HTTP API). The app lives in its own repo; this module is only the DEPLOY, pinned by the signal-sift flake input.

Why a user service, from the app's own moduleCopy Anchor Link

• The daemon needs no privilege: it reads public content and writes under my XDG directories. So it is a systemd user unit started with my session, not a system service.
• The unit comes from the app's Home Manager module (homeManagerModules.default), not from a unit written here: the hardening, Nice=10, CPU/IO weight 20 and MemoryMax=10G are decided and measured in the app's repo (its deploy/nixos/README.md and ADRs 0008 and 0011).
• No inputs.nixpkgs.follows on the input: the package and the Vulkan builds of whisper.cpp and llama.cpp come from SignalSift's own lock, the exact store paths it benchmarked, even though this repo is on nixos-26.05.

GPU onlyCopy Anchor Link

Speech recognition and the LLM run on the Arc B580 or not at all (runtime.require_gpu, stated in the module even though it is the default). Both engines silently fall back to the CPU when Vulkan fails; the app kills such a run and retries the job later instead. MEASURED on 15/10/2026 on this machine: CPU whisper is about 11 times slower than Vulkan and keeps all six cores busy. ffmpeg and yt-dlp stay on the CPU, since decoding audio and downloading are not inference.

The models are declaredCopy Anchor Link

The three model files are fetchurls with their SHA-256, the same files the app's benchmarks used, so a reinstall needs no manual download (rule 3). On 16/10/2026 they were added to the store from the copies already on disk (nix-store --add-fixed sha256), so the first build fetched nothing. The other models in ~/.local/share/signalsift/models were only for the benchmarks.

The interest profile is hereCopy Anchor Link

signalsift-interests.toml is versioned in this repo, so a change to what I find relevant has a commit. Only analyses made after an edit use the new text; the app stores a fingerprint of the profile with every score.

Not declared: the sourcesCopy Anchor Link

The feeds and channels live in the app's database, not in Nix, because poll state, history and my feedback are attached to them (rule 6). After the first switch, once:

signalsift doctor
signalsift sources import ~/Projects/GitHub/v1cferr/signal-sift/config/sources.example.toml

RejectedCopy Anchor Link

• A unit written here: it would duplicate the hardening and limits the app already owns and measures (rule 14).
• Following this repo's nixpkgs: nixos-26.05 would rebuild the engines at versions nobody benchmarked.
• Models downloaded by hand: a manual step a reinstall would forget (rule 3).