8000 farolanf (basysyar) · GitHub
[go: up one dir, main page]

Skip to content
View farolanf's full-sized avatar

Block or report farolanf

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
farolanf/README.md

farolanf

I build small SaaS products, and the machinery that decides which one gets worked on next.

Mostly I'm not hand-coding any more. I design the architecture and the guardrails, then unattended agent sessions do the mechanical work, and nothing counts as done until it's verified against real evidence. That green wall below is almost all private repos, and most of it is agent-authored. The volume isn't the interesting part. What's interesting is what the guardrails refuse.

fleet stats

How I work

  • Agents work here unattended, and when something needs my judgment they DM me with the options, so I either tap one or type a direction they didn't offer. We build it together from there, and once it ships it runs its own loop: production errors get filed, fixed and deployed without me in the middle of it, and I come back in only where there's a call to make. Destroying data is the one thing they never do alone.

  • Coverage is a smoke detector, and one repo-wide number lies in both directions. One product here reads 66.91% overall, which tells you nothing: the domain is gated at 99 and reaches 99.22, the adapters sit at 0 with the reason written beside the exclusion, and each gate is set a little under what the suite actually reaches so it fires on a regression instead of on noise. I started doing it this way after finding 300 passing tests, an architecture guard, and the two rules the product exists for at 13%.

  • Reading a whole file to find one thing is the expensive habit, so I built a grep-able digest of every symbol and the line it sits on. One model file here costs about 26,000 tokens to read, its digest costs 2,900, and the answer to "what runs when this record saves" costs 140. There's no daemon and nothing to remember to re-run: it reconciles against the working tree on every call, because a copy that syncs on a timer is a second source of truth, and between the edit and the sync it answers confidently and wrongly.

  • In this loop the products harden over time on their own: an incident produces a rule, the rule gets a test, and every run after that carries it. What's left for me isn't writing the code any more, it's noticing when a hardening stopped at the product it happened in and never travelled, and then making that noticing automatic too.

Stack

TypeScript on Cloudflare edge, Rust for the rigorous core, Rails over deep PostgreSQL, Python for the LLM services. Production SaaS with the un-fun correctness in it: billing, credits, subscriptions, webhooks that fan out.

Popular repositories Loading

  1. DragDropInterface DragDropInterface Public

    Forked from StackHive/DragDropInterface

    A Cross frame drag and drop interface for web browsers

    JavaScript 1

  2. juwalnode juwalnode Public

    JavaScript 1

  3. json-log-viewer json-log-viewer Public

    Forked from gistia/json-log-viewer

    Powerful terminal based viewer for JSON logs using ncurses.

    JavaScript 1

  4. webhook-listener webhook-listener Public

    Run commands on webhook event

    TypeScript 1

  5. jops jops Public

    Automatically exported from code.google.com/p/jops

    Java

  6. diaspora diaspora Public

    Forked from diaspora/diaspora

    A privacy-aware, distributed, open source social network.

    Ruby

0