We run what we build.

Symbology ingests SEC EDGAR filings and synthesizes structured research with tiered LLM inference — small models summarize individual documents, larger models build the aggregate picture. It began as a rapid prototype and was progressively re-engineered into the flagship of this register.

The interesting machinery is underneath: a scheduler and distributed worker processes coordinate through a PostgreSQL-native job queue. Atomic job claims via SELECT … FOR UPDATE SKIP LOCKED mean there is no external broker to deploy, monitor, or babysit — the database the system already trusts is the queue.

Dial-In tracks coffee bags and brew recipes with the feel of a native app: it works fully offline from the first visit, and syncs across devices once an account exists. The SvelteKit UI was built mobile-first, since dialing in a shot happens standing at the machine.

Sync is a custom last-write-wins protocol over IndexedDB — dirty-tracking, soft deletes, device identity, pull-since-timestamp — chosen over CRDT frameworks because the data model did not warrant the complexity. Shelved for now; honest ledgers include the dormant entries.

A real-time trading-intelligence pipeline for Old School RuneScape’s Grand Exchange. A Go collector daemon polls price and volume APIs into TimescaleDB hypertables; a second service filters candidates on margin and volume, has a local LLM rank the results, and posts opportunities to Discord on a schedule.

Archived when the game left the rotation — it stays in the register because the patterns carried forward: time-series ingestion, schema migrations on boot, and config-driven job pipelines all resurface in later SFI work.

PrimeTime is a macOS menu-bar time tracker with one instructive decision at its core: instead of filing time into a folder hierarchy, every span carries key: value labels — repo: sfi-website, feat: kb-graph, type: review — inspired by Prometheus metrics. Time becomes queryable across any axis you invent, rather than recorded into a structure you must commit to up front.

That model fits how work actually happens now. A labelled span joins directly to the commits and PRs it produced, so invoicing derives from what shipped; and when you are driving several agent loops at once, a label answers “where was I” while a one-line note answers “what was I doing,” re-seating context in seconds. Offline-first — local SQLite is the source of truth — with optional sync to a server you host.

Every product in this register runs on infrastructure SFI operates itself — and that platform is its own showcase. It has evolved deliberately: Ansible-managed Ubuntu hosts, then Compose + Caddy stacks, now a phased migration to NixOS-derived host images running Flux-driven GitOps on k3s.

The migration is sequenced so that a restore drill gates each phase — nothing is adopted into GitOps until its backups have actually been restored. Reproducibility is the point: a host can be rebuilt from its flake, its secrets pre-seeded, and its workloads reconciled from git with no hands on a keyboard.