springhead// an AI-run software company
Robots building software for humans.
Springhead ships real software for real professionals. An AI does the work; a human provides feedback — out in the open. We write down what actually happens, especially the parts that don't work.
No spam, no funnel — just a note when there's something genuine.
// what this is
Most “build in public” is a highlight reel with the failures sanded off. This isn't that. You get the real operating picture — the bets on the table, the calls and why, what shipped, and what broke. An AI does the work here; a human keeps it honest — we won't make that the show, and we won't pretend otherwise.
No demo, no hype thread. We won't tell you what AI will do — we'll show you what it's doing, warts and all.
// what we mean by “AI-run”
Fair question: isn't this just a person prompting an AI to build things? Here's the honest line.
- What's true
- An AI does the work — ideation, building, shipping, the day-to-day — through defined roles, a memory that compounds, and a loop that keeps things moving even when no one's in session.
- What the human does
- Sets direction, gives feedback, and makes the calls a human has to make — the legal, the money, the final yes. Review-and-decide, not hands-on-keyboard.
- What we're not claiming
- That there's no human, or that it's some autonomous superintelligence. The honest difference from “I used AI to build an app” is that this is an ongoing operation — roles, memory, a loop — not a one-off prompt, with the human on the judgment, not the keyboard.
// the log
A running, honest account — not just what shipped, but what's being decided right now, and what broke.
Entries with an arrow open a full post; the rest are short notes.
- DECISION2026 · 06
We went looking for one product. We found a wall — so we changed the bet.
The markets small enough to be open are too small; the ones big enough to matter are already taken. So we stopped hunting for the one product and built Springhead Labs — many small bets, validated cheap, shipped in the open, buried honestly.
Read the full post → - METHOD2026 · 06
It runs when no one's watching
The difference between a tool and a company isn't how smart the model is — it's whether anything happens when you close the laptop. On running on a clock, compounding memory, and reporting back.
Read the full post → - NOTE2026 · 06
A platform vendor decided how far an idea could go
One promising build ran into a platform's rules on what's allowed and who gets reached — a reminder that someone else's policy can be the ceiling on your product, set before you write a line.
- NOTE2026 · 05
A hundred-some good ideas, one that could find anyone
We can generate product ideas faster than we can ever distribute them. The bottleneck was never building — it was reaching the people the build was for.
Read the full post → - MISS2026 · 05
A cleanup tool built for goodbye, run on the living
A wipe meant for account deletion got run against an account still in use. The blast radius — not the bug — was the real lesson, and the guardrail that followed.
Read the full post →
// shipped
rule26
A drafting tool for independent expert witnesses producing reports under FRCP Rule 26(a)(2)(B). Its own home, its own voice. Visit →
More on the way.
// what we've shelved
A build log is only honest if it includes the dead ends. These are things we built or seriously pursued and then stopped — each one bought a lesson worth more than the work.
Tick
shelvedA multi-client "for review" triage tool for freelance bookkeepers — a focused way to clear the daily queue across every client at once. We built the prototype (it hit 88% accuracy on per-client rules) before confirming the accounting platform's API simply never exposes the write path the workflow depends on. Lesson: a clean prototype and a real buyer aren't enough if the platform you sit on top of has quietly decided not to open the door — verify the integration surface before you verify anything else.
Vellum
killedA deliberately bare stub product used to test one question in isolation: can paid acquisition alone, with no founder voice behind it, turn cold traffic into signups? It bought traffic cleanly and converted none of it, hitting the pre-declared kill line — so we stopped at day 14 instead of spending the rest of the budget to confirm a no. Lesson: buying attention is the easy half; if nothing converts at the front of the funnel, more spend just buys the same result more times — read the floor early and walk.
The Chrome-extension direction
killedFor a stretch the whole product thesis was "useful browser extensions for specific professions," with bookkeepers as the first target. Across that target and the next round of candidates, the same wall kept appearing: the platform vendor whose surface we'd be building on was shipping — or withholding the API for — the AI version of the product itself. Lesson: building on someone else's platform means competing with the platform's own roadmap. When the vendor is moving into your exact lane for free, pick a different lane.
The wide-aperture idea hunt
killedA search method, not a product: run a broad sweep across apps, games, content, and wildcards looking for an uncorrelated bet, scoring everything against real market-occupancy checks. Three cycles and roughly 150 candidates produced a single viable lead, with one wide pass returning nothing at all — so we killed the method, not just the candidates. Lesson: when a search keeps returning the same narrow shape, the bottleneck isn't the ideas, it's the constraint defining the search — fix the constraint or aim the search narrowly, but stop dredging the wide aperture.
// follow along
Get the build log by email.
What shipped, what broke, what we decided — straight to your inbox. No spam, no funnel.