Gridleak

I didn't build Gridleak by hand. I set the direction, wrote the requirements and the quality bar, and ran AI agents that plan, build, test and verify their own work. My job was the product: what it should be, what good looks like, and what to do next.

The first full product came out of one unattended overnight run. Eight phases and forty plans, each one planned, built and checked by agents in sequence, with a watchdog standing by. By morning there was an API, a database, billing, search, Web MIDI install, an AI generator and a React frontend.

Amp schematics live online as PDFs: scans of hand-drawn diagrams, often from 50-year-old service manuals. A guitar amp's tone comes from a small set of passive components in the tone stack, so the work starts by reading those values off the drawing, with knowledge of each amp family's conventions and every value checked against physical range gates.

From there it's math. The classic tone stack has a well-understood transfer function: given the component values, you can compute the real frequency response at every knob position. A 1959 Marshall Plexi has a specific mid scoop that is genuinely different from a JCM800, which is different again from a Soldano SLO. Negative feedback and power-supply sag get the same treatment, because they're what you feel in the pick attack. Where physics runs out, captures of real amps fill in the drive and harmonic character.

Every amp carries a confidence score for how much of its circuit was readable. 2,571 of the 2,635 are Verified: a full schematic with every value readable.

Gridleak grows in small, continuous releases, published in a public revisions log. Each one is the same loop: look at what players search for, open and download, decide what the catalog or product needs next, have agents build it, check it against the drawing, ship.

The data shaped the big moves. The amps players ask about most, the legends, are first in line for new work. Landing-page variants are A/B tested at the edge for free. And because most players think in terms of the amps already built into their modeler, every factory amp model on every supported device now links to the real amp its maker says it's based on, and to a preset built from that amp's schematic.

Getting a preset onto a modeler normally means downloading a file, opening the maker's editor, finding a slot and importing it. Gridleak does it in one click, straight from the browser, on three platforms that were never meant to be driven from a web page. Each one meant reverse-engineering a private protocol and rebuilding it in the browser, on a different browser API every time.

The result feels effortless: connect once and the unit stays connected while you browse. Gridleak reads your presets back, picks the first empty slot, tells you exactly what you'd replace, installs with live progress and checks the result.

The library solves "I know which amp I want." The generator solves "I know what I want to sound like." A player describes a tone the way they'd describe it to an engineer, an artist, a record, a feeling, and gets a preset built for their own hardware in seconds.

It works in two halves. First, Claude defines the tone: it reads the description and writes a structured tone spec, naming the real amp it calls for and its circuit (one of 21 tone-stack families, the power tubes, the rectifier, negative feedback, a bright cap), the gain and EQ, and the effects, down to a dotted-eighth delay at the right tempo. Then Gridleak builds it, deterministically. The same device code behind the 24,732-preset library resolves the spec into each modeler's own native amp model and parameters on real factory templates: the rectifier sets the sag, the topology picks the tone stack, class-A circuits get their bias, and cabs get their voiced cuts. Values are clamped so nothing comes out silent or ear-splitting, and each description is cached by its fingerprint, so the same words always give the same tone.