Machine learning engineers could prototype incredible workflows in Jupyter notebooks, but translating them to production-ready APIs took months. The gap between "this works on my machine" and "this serves 10,000 requests/day" was killing innovation cycles at enterprises.

While PlaiDay was burning thousands a day on compute costs, three conversations happened in the same week:

1. An enterprise client asked if they could license our backend instead of our consumer app

2. A ML engineer showed us a 500-line file just to run a modified Stable Diffusion

3. Our own team was shipping ML features 10x faster than any company we knew

The Insight: Our competitive advantage wasn't our consumer features—it was our ability to transform ML research code into production services in under 4 weeks. We'd accidentally built what every ML team needed.

I led 100+ developer interviews in 14 days. Instead of asking "would you use this?", I asked:

• "Show me your last deployment. Walk me through every step."

• "What percentage of your sprint is actual ML work vs. DevOps?"

• "How many times have you rebuilt the same inference pipeline?"

More than 70% of ML engineers were rebuilding the same workflow orchestration logic at every company. The problem wasn't access to models—it was productionizing them.

• Workflow orchestration engine (handled 50K+ executions/day)

• Model serving infrastructure

• Queue management system

• Social features (removed 50K lines of code)

• Mobile apps

• Consumer authentication system

• Cloud-based visual workflow builder (core product, modified ComfyUI)

• Custom drag and drop node system built on ReactFlow

• API key management

• Usage-based billing

• Workflow versioning

The controversial bet: Going all-in on visual workflow building. While every other tool forced developers to write configuration files or code, we believed ML engineers needed to see their pipelines – all visual, all in the cloud, no local setup required.

The breakthrough: Cloud-native visual ML pipelines. Drag a Stable Diffusion node, connect it to an upscaler, add a background remover—deploy to production in 30 seconds. No code, no Docker, no GPU provisioning.

While competitors forced developers to write Python or YAML, we bet everything on visual workflow composition. ML engineers could finally see their data flowing through models in real-time.

This became our wedge. Enterprises had million-dollar MLOps platforms requiring weeks of setup, but developers were bypassing them to prototype and deploy on Salt in minutes.

Instead of a traditional product hunt launch, we hosted 50+ ML engineers in our LA office:

• Live-coded a computer vision pipeline in 10 minutes

• Had several open-source maintainers present their workflows

• Collected 1000+ beta signups in one evening

Open source community would drive enterprise adoption

Open source users generated 90% of support tickets, 5% of revenue.

• 60,000 signups → 40,000 activated users (created a workflow)

• 40,000 activated → 10,000 adopted users (developing workflows weekly)

• 10,000 adopted → 1,500 revenue generating developers

• Industry average: 3-4 weeks

• Salt AI average: 3-4 days

We spent a lot our time supporting free users who were never going to pay. Every open source user had specific preferences. Meanwhile, enterprise customers were begging to pay us $50K/year for basic features we hadn't built yet.

The best B2B products come from solving your own problem at scale. We didn't set out to build developer tools—we just packaged our survival mechanism from PlaiDay and discovered hundreds of teams needed the same thing.

Aber Whitcomb, Chris DeWolfe, Jim Benedetto, Charlie Basil, Scott Baggett, Alex Duffy, Ryan Wang, Mat Blysz