True Assistant

ON ICE

Closed-source company AI platform that keeps data off external model providers, records meetings with transcripts, and answers questions about company decisions in real time.

date
tags
automation, agents
stack
n8nTypeScriptAI
context
Company-wide meeting ops and grounded AI assistant — invite-only pilot
links
liverepo

Test access only. True Assistant is in private beta. Sign-in at trueassistant.piyushtater.com is for approved testers. If you are not on the list yet, apply below.

Apply for test access

Use the Request beta access link (also on the project card in the graph). Your email app will open with a short template — fill in:

  • Your full name
  • Your company or organization name
  • Your position / role in that organization
  • What you want to know or try, and why you need access
  • Any details relevant to evaluating your request
  • The Google account email access should be granted on (used for sign-in)

I review applications manually and reply within a few business days. If approved, I email you when your account can sign in.

Prefer a form? A Google Form intake is being set up; until then, the email link above is the application path. Fallback: piyushtater158@gmail.com with subject True Assistant beta access request.

What it does

True Assistant (also shown in demos as Meeting-Ops / Kaira) is a closed-source company AI platform. Because it is closed source and self-hosted on infrastructure you control, the company does not need to share data with external model providers. Use it company-wide for AI-driven work: every meeting is recorded and transcribed, and an assistant can answer questions about the company database or past decisions in real time — with human approval before agentic actions go live.

How it works

  1. Sign in with a Google work account (invited pilot users only).
  2. Capture — a bot joins calls, transcribes, and stores meeting memory.
  3. Extract + trust gate — dual-pass extraction of decisions and action items; approved items sync to tools (e.g. Slack) with an audit trail.
  4. Ask — grounded Q&A across meetings (“what did we decide about X in January?”).
  5. Approve — routine agent actions (scheduling, follow-ups) stay behind a human-in-the-loop UI.

What I built

  • Meeting-Ops review UI on Vercel with Google OAuth for pilot users
  • n8n orchestration on a controlled VPS for capture, extraction, sync, and Q&A
  • Hybrid memory in Supabase (Postgres + pgvector) for cross-meeting recall
  • Human approval path so agentic writes never ship without a person in the loop
  • Closed-source deploy path so company data stays on infrastructure you control

Stack

n8n · TypeScript · Supabase · self-hosted AI · Vercel