Private AI tools that are new on GitHub and not yet on anyone's shopping list — we test them, package them and set them up for small teams in the US and Australia. Each one is a pilot: honest about what works today.
meeting.m4a · 42 min · stays on your server
Transcript ready · 3 speakers · local Whisper
What did we promise Acme on delivery?
✓ “Two weeks after sign-off” · at 18:42 in the call
No audio sent to the cloud“Our meetings are confidential. We can't use cloud note-takers.”
Law, health and HR calls shouldn't be uploaded to a stranger's servers to get a summary.
“The answer is in our database. I just can't ask it.”
Owners wait days for a report that a well-asked question could answer in seconds.
“Our AI forgets everything between chats.”
Assistants that don't remember customers or past decisions repeat the same questions forever.
We watch the fastest-growing open-source AI projects, try them hard, and only pilot the ones that hold up.
We start from your problem — confidential meetings, slow reports, forgetful assistants — not from the tool.
We pick the open-source project that fits, check its licence and maturity, and agree what a pilot must prove.
A small pilot on your own server with local AI, so nothing confidential leaves your network.
Two to four weeks of real use, a written result, then a decision: keep, change or drop.
Private AI tools merged from the most promising new open-source projects — offered as pilots.
Project 01
Private meeting notes: record, transcribe and summarise on your server with local AI, then ask what was promised — with a timestamp.
Project 02
Text-to-SQL over your database and CRM in plain English, with every query shown for review and AI costs kept under control.
Project 03
AI memory for support and sales assistants: they remember each customer's history and past decisions, safely.
Project 04
A small team's private AI setup: safe sandboxed agents, passkey logins and connections to your own tools through MCP.
Project 05
An MCP server that lets ChatGPT, Claude or your own agent safely read and act on your CRM, ERP or database.
Project 06
Thousands of old PDFs and scans turned into an AI-ready knowledge base that gives deep, cited answers.
New tools change fast. We check licences, run them privately and say clearly what's still a pilot.
Some new projects have custom licences. We read them before recommending anything.
Pilots run on models you host, so confidential information never leaves your network.
Every pilot has a success measure and a written result. If it doesn't help, we say so.
We'd love AI meeting notes, but client calls can't go to a cloud service.
Then we'll run transcription and summaries on a machine in your office. Let's pilot it on internal meetings for two weeks first.
And if it isn't good enough?
Then the pilot report says so and you've lost two weeks, not a year's subscription. That's the point of a pilot.
An example of a typical first call.
Free · 30 min
Your problem, the candidate tools, and what a pilot would need to prove.
Best when you're curious but cautious.
2–4 weeks
A private setup used for real work, with a written result at the end.
Best for testing before committing.
After the pilot
Hardening, backups, monitoring and support for the tools that earned it.
Best once the pilot proves its value.
We'll reply within one business day with what's worth piloting — and what isn't.