Bart Labs — The AI Product Lab

What we believe

I build the tools I needed and couldn't find — or found priced five times what they were worth, shipping half of what they promised. Every product here started as my own working problem.

The through-line is evidence. Software should show its reasoning, keep a record you can inspect, and leave judgment with the person who has to answer for the outcome. That principle scales from a compliance team down to a photographer culling a wedding.

Michael Bartels · Founder, Bart Labs

What we build

One discipline. Three domains.

Enterprise · Early Access

ForgeShield

One control layer for every AI vendor

Your teams buy AI from several vendors at once. ForgeShield sits in front of all of them — one set of rules on what can be sent, one record that can't be altered afterwards, one memory that survives changing vendor. It assists compliance; it never claims to make you compliant.

Explore ForgeShield → Photographers · Available

LrForge

Evidence-aware analysis for Lightroom

An assistant that shows why a frame works — light, composition, focus — grounded in what it actually sees. The photographer keeps aesthetic authority, always.

Explore LrForge → One-time purchase · runs local Creative · Beta

Storygild

Evidence-aware writing

Long-form writing where every suggestion traces back to your notes and the author revises on top — source, suggestion, decision, final passage, all kept in the thread.

Explore Storygild →

Why "the memory & evidence layer"

Measured, not asserted.

⬢

AI vendor adapters governed by one ForgeShield policy set

⛓

1.3M

sealed records in the ForgeShield audit chain, verifiable without trusting us

◇

3,500+

requests in the LrForge field study — measured where AI helps and where it doesn't

⬡

Local

both run on hardware you own — nothing has to reach a cloud server

ForgeShield figures measured on our own production instance, 27 August 2026. ForgeShield assists regulated teams in demonstrating and investigating AI use. It supports compliance workflows; it does not make an organization compliant.

Field papers

The lab publishes its work.

Can AI tell a good photograph?

Where AI helps a photographer — and where the human must stay in charge

A 3,500-request study of local-model image analysis: where an evidence-aware assistant adds signal to culling and editing decisions, and where aesthetic authority has to remain human.

LrForge · 3,500+ requests · ~12 min read

What does an AI evidence record look like?

Reconstructing a single AI interaction into a tamper-evident record

Input → PII removed → model response → policy decision → retention → Merkle-linked proof. How ForgeShield turns one prompt into something a regulated team can investigate later.

ForgeShield · architecture note · ~8 min read

Evaluating AI governance?

See how ForgeShield builds an evidence trail your auditors can verify. No sales gauntlet — a real walkthrough.

Request a walkthrough →

Working in Lightroom?

Evidence-aware analysis that respects your eye. One-time purchase, runs on your own machine.

Explore LrForge →