DayStorm Labs

From Meeting to Merge.
Autonomous.

Voice notes become specs.
Specs become merged PRs.
No human in the loop.

Production · weeks live · 880+ builds · 666 merged
experiment — pr #174 · live receipt
input       voice note (52s, English)
extract     meeting item · confidence 87/100
spec        two-phase generator · 14 acceptance criteria
build       factory autonomous · 14-step pipeline · self-verifying
duration    36 minutes
human equiv ~8 hours
result      MERGED · 14 of 14 acceptance criteria green
// no human review touched the build
The Stack

Two systems. One closed loop.

The customer-facing stack is two composable layers. Together they form the loop from raw intent to merged code — and back to the next intent.

Distillery
Front
Where meetings become specs.
Multi-tenant managed service. Meetings and voice notes go in. Items, signals, and acceptance criteria come out — gated by an 8-dimension precision framework before anything reaches the build.
  • Voice-note-to-spec pipeline live in production
  • Plenty of features already shipped through the loop
  • Currently onboarding select customer projects
Factory
Back
Where specs become merged code.
A 14-step autonomous pipeline. Pre-filter, build, judge, visual review, E2E generation, healing loops. Auto-merges when the satisfaction score clears the gate. Self-building.
  • 880+ autonomous builds across 10 configured repos
  • 666 features delivered, ten weeks production
  • Three architecture iterations in those weeks
How it works

One loop. Five beats. Closed.

record extract generate build merge
next meeting picks up where this one shipped

Voice or meeting in. Distillery extracts intent, scores confidence, generates the spec.

Factory's 14-step pipeline ships it autonomously, with self-verification at every gate.

The merged feature shows up where it was meant to live. The next meeting, the next voice note, picks up from there. The loop closes through use, not through reflection.

Why the specs land

Distillery doesn't write specs in a vacuum. Before a single acceptance criterion is generated, it has the full Product Vision Document and the live codebase as context. Every spec is grounded in what you're building and in what already exists.

It's like having the senior engineer, the product lead, and the codebase architect all at the same table — triggered by a 30-second voice note from one person.

Empirical findings

Measured, not promised.

Finding 01
Spec precision is the largest quality lever.
Same spec, same model, same context. Only the acceptance criteria wording changed. Judge score went from 80% to 98%. The interpretation room you remove from the input is the output you no longer have to fix.
Finding 02
Recency beats embedding for healer memory.
Last 5 failures per repo, injected into the healer prompt as plain text. No vector lookup. Most repeated bugs converge on retry 1 instead of retry 2.
Finding 03
Spec-first cuts AI cost roughly in half.
Once confidence assessment runs before the question loop, per-item AI cost in the spec-generation pipeline drops by about 60%. The cheapest spec is the one you don't ask unnecessary questions about.
The thesis

Where this goes.

Solo operators won't scale by hiring. They'll scale by building infrastructure that ships while they sleep.

The bottleneck is no longer code production. It's spec quality and feedback-loop closure. The systems running here are the working hypothesis: a one-person product company on three composable layers — one for capturing intent, one for shipping code, one for accumulating wisdom.

Currently onboarding the first wave of customers. Solo operators and small teams who want the same leverage.
Side project

DarkFactory: The Book.

A live-written chronicle of building Daystorm Labs and the systems that ship its software.

Every night, a harvest pipeline reads the day's Claude sessions and daily notes. The relevant fragments are distilled into structured nuggets. Once a week, those nuggets are synthesized into a narrative chapter. The book is being written by the system it describes — with me as the editor.

Status 11+ weekly chapters live · published when it's ready
Connect

If you're building or want to use this, let's talk.

Hands-on onboarding, hands-off operation. Customers run their own projects on the Distillery once their workspace is set up. I operate the Factory backend.
Want to use the Distillery for your team
Onboarding inquiry — let's see if there's fit.
What I'll need project context, GitHub repo access, and your team setup
Building something similar
Compare notes, swap patterns.
What I'll bring live URLs, dashboards, and the actual receipts — not slides
Investing or partnering in this space
Intro for a longer conversation.
What I'll send first a short writeup of where this stands and where it's heading

Just following what's next? LinkedIn — I post when something ships.

🇺🇸 Heading to Code with Claude in SF?

Find me at the event or DM via LinkedIn. Office hours, coffee, after-party — happy to swap receipts in person.